lightbox layout and x close

Co-authored-by: Michael Chung <michaelchung@michaels-mbp-2.home>

.gitignore

@@ -66,3 +66,7 @@ scraps/
 # These should be built by the platform
 windows/sqlite/sqlite3.exp
 windows/sqlite/sqlite3.lib
+
+# ArtCraft Windows DLLs
+crates/desktop/src-tauri/*.dll
+windows_dlls/

Cargo.toml

@@ -1,6 +1,8 @@
 [workspace]
 resolver = "2"
 
+# This block defines the workspace members.
+# When we add a library or binary crate, it needs an entry here.
 members = [
   # Types: common strongly-typed tokens, ids, payloads, etc.
   "crates/_types/composite_identifiers",
@@ -22,7 +24,10 @@ members = [
   "crates/cli/production/migration-tool",
 
   # Desktop apps
-  "crates/desktop/src-tauri",
+  "crates/desktop/ml_models",
+  "crates/desktop/ml_weights_registry",
+  "crates/desktop/tauri-4o",
+  "crates/desktop/tauri-realtime",
 
   # Shared libraries
   "crates/lib/actix_cors_configs",
@@ -30,6 +35,7 @@ members = [
   "crates/lib/api_clients/cloud_storage",
   "crates/lib/api_clients/google_drive_common",
   "crates/lib/api_clients/google_sign_in",
+  "crates/lib/api_clients/openai_sora_client",
   "crates/lib/api_clients/thumbnail_generator",
   "crates/lib/bootstrap",
   "crates/lib/bucket_paths",
@@ -56,6 +62,7 @@ members = [
   "crates/lib/files/mimetypes",
   "crates/lib/files/videos",
   "crates/lib/idempotency",
+  "crates/lib/image_algorithms",
   "crates/lib/jobs_common",
   "crates/lib/logging",
   "crates/lib/markdown",
@@ -115,3 +122,35 @@ members = [
   # see usage: https://crates.io/crates/cargo-hakari
   "crates/my-workspace-hack",
 ]
+
+[workspace.package]
+# Common inheritable settings for monorepo packages
+publish = false
+
+[workspace.dependencies]
+# Hakari workspace hack (faster compile times)
+my-workspace-hack = { path = "crates/my-workspace-hack" }
+
+# External libs
+# Centrally pinned versions so that we can uniformly upgrade/downgrade versions across the workspace
+# If a crate needs to use an older/newer version, the dependency should be managed in that crate's own Cargo.toml
+anyhow = "1.0.97"
+chrono = "0.4.40"
+log = "0.4.27"
+once_cell = "1.21.3"
+rand = "0.9.0"
+reqwest = { version = "0.12.15", features = ["json", "cookies"] }
+serde = {  version = "1.0.219", features = ["derive"] }
+serde_derive = "1.0.219"
+serde_json = "1.0.140"
+serde_yaml = "0.9.1"
+toml = "0.7.4"
+
+# Internal libs
+crockford = { path = "crates/lib/data/crockford" }
+errors = { path = "crates/lib/errors" }
+filesys = { path = "crates/lib/files/filesys" }
+memory_caching = { path = "crates/lib/caching/memory_caching" }
+openai_sora_client = { path = "crates/lib/api_clients/openai_sora_client" }
+primitives = { path = "crates/lib/data/primitives" }
+subprocess_common = { path = "crates/lib/subprocess_common" }

app_config.yaml

@@ -1,7 +1,8 @@
-image_width: 512
-image_height: 512
+image_width: 1024
+image_height: 1024
 #scheduler_steps: 15
 # NB(bt): It looks like we can get down to 2 sampling steps, perhaps ?
-scheduler_steps: 4
+scheduler_steps: 8
 scheduler_samples: 1
-cfg_scale: 1.1
+cfg_scale: 15.0
+seed: 42

crates/_types/enums/src/by_table/generic_inference_jobs/inference_job_type.rs

@@ -40,6 +40,10 @@ pub enum InferenceJobType {
   #[serde(rename = "studio_gen2")]
   StudioGen2,
 
+  /// Sora GPT 4o image gen
+  #[serde(rename = "image_gen_api")]
+  ImageGenApi,
+
   /// A job that turns "FBX" game engine files into "GLTF" files (Bevy-compatible).
   #[serde(rename = "convert_fbx_gltf")]
   ConvertFbxToGltf,
@@ -108,6 +112,7 @@ impl InferenceJobType {
       Self::GptSovits => "gpt_sovits",
       Self::ComfyUi => "comfy_ui",
       Self::StudioGen2 => "studio_gen2",
+      Self::ImageGenApi => "image_gen_api",
       Self::ConvertFbxToGltf => "convert_fbx_gltf",
       Self::MocapNet => "mocap_net",
       Self::RvcV2 => "rvc_v2",
@@ -132,6 +137,7 @@ impl InferenceJobType {
       "gpt_sovits" => Ok(Self::GptSovits),
       "comfy_ui" => Ok(Self::ComfyUi),
       "studio_gen2" => Ok(Self::StudioGen2),
+      "image_gen_api" => Ok(Self::ImageGenApi),
       "convert_fbx_gltf" => Ok(Self::ConvertFbxToGltf),
       "mocap_net" => Ok(Self::MocapNet),
       "rvc_v2" => Ok(Self::RvcV2),
@@ -159,6 +165,7 @@ impl InferenceJobType {
       Self::GptSovits,
       Self::ComfyUi,
       Self::StudioGen2,
+      Self::ImageGenApi,
       Self::ConvertFbxToGltf,
       Self::MocapNet,
       Self::RvcV2,
@@ -197,6 +204,7 @@ mod tests {
       assert_serialization(InferenceJobType::GptSovits, "gpt_sovits");
       assert_serialization(InferenceJobType::ComfyUi, "comfy_ui");
       assert_serialization(InferenceJobType::StudioGen2, "studio_gen2");
+      assert_serialization(InferenceJobType::ImageGenApi, "image_gen_api");
       assert_serialization(InferenceJobType::ConvertFbxToGltf, "convert_fbx_gltf");
       assert_serialization(InferenceJobType::MocapNet, "mocap_net");
       assert_serialization(InferenceJobType::RvcV2, "rvc_v2");
@@ -220,6 +228,7 @@ mod tests {
       assert_eq!(InferenceJobType::GptSovits.to_str(), "gpt_sovits");
       assert_eq!(InferenceJobType::ComfyUi.to_str(), "comfy_ui");
       assert_eq!(InferenceJobType::StudioGen2.to_str(), "studio_gen2");
+      assert_eq!(InferenceJobType::ImageGenApi.to_str(), "image_gen_api");
       assert_eq!(InferenceJobType::ConvertFbxToGltf.to_str(), "convert_fbx_gltf");
       assert_eq!(InferenceJobType::MocapNet.to_str(), "mocap_net");
       assert_eq!(InferenceJobType::RvcV2.to_str(), "rvc_v2");
@@ -243,6 +252,7 @@ mod tests {
       assert_eq!(InferenceJobType::from_str("gpt_sovits").unwrap(), InferenceJobType::GptSovits);
       assert_eq!(InferenceJobType::from_str("comfy_ui").unwrap(), InferenceJobType::ComfyUi);
       assert_eq!(InferenceJobType::from_str("studio_gen2").unwrap(), InferenceJobType::StudioGen2);
+      assert_eq!(InferenceJobType::from_str("image_gen_api").unwrap(), InferenceJobType::ImageGenApi);
       assert_eq!(InferenceJobType::from_str("convert_fbx_gltf").unwrap(), InferenceJobType::ConvertFbxToGltf);
       assert_eq!(InferenceJobType::from_str("mocap_net").unwrap(), InferenceJobType::MocapNet);
       assert_eq!(InferenceJobType::from_str("rvc_v2").unwrap(), InferenceJobType::RvcV2);
@@ -261,7 +271,7 @@ mod tests {
     fn all_variants() {
       // Static check
       let mut variants = InferenceJobType::all_variants();
-      assert_eq!(variants.len(), 16);
+      assert_eq!(variants.len(), 17);
       assert_eq!(variants.pop_first(), Some(InferenceJobType::VideoRender));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::LivePortrait));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::FaceFusion));
@@ -269,6 +279,7 @@ mod tests {
       assert_eq!(variants.pop_first(), Some(InferenceJobType::GptSovits));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::ComfyUi));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::StudioGen2));
+      assert_eq!(variants.pop_first(), Some(InferenceJobType::ImageGenApi));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::ConvertFbxToGltf));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::MocapNet));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::RvcV2));

crates/_types/enums/src/by_table/generic_inference_jobs/inference_model_type.rs

@@ -33,6 +33,10 @@ pub enum InferenceModelType {
   #[serde(rename = "seed_vc")]
   SeedVc,
 
+  /// NB: This is for Sora GPT 4o image gen
+  #[serde(rename = "image_gen_api")]
+  ImageGenApi,
+
   #[serde(rename = "tacotron2")]
   Tacotron2,
   #[serde(rename = "vits")]
@@ -70,6 +74,7 @@ impl InferenceModelType {
       Self::VallEX => "vall_e_x",
       Self::RerenderAVideo => "rerender_a_video",
       Self::StableDiffusion => "stable_diffusion",
+      Self::ImageGenApi => "image_gen_api",
       Self::SeedVc => "seed_vc",
       Self::MocapNet => "mocap_net",
       Self::StyleTTS2 => "styletts2",
@@ -90,6 +95,7 @@ impl InferenceModelType {
       "vall_e_x" => Ok(Self::VallEX),
       "rerender_a_video" => Ok(Self::RerenderAVideo),
       "stable_diffusion" => Ok(Self::StableDiffusion),
+      "image_gen_api" => Ok(Self::ImageGenApi),
       "mocap_net" => Ok(Self::MocapNet),
       "styletts2" => Ok(Self::StyleTTS2),
       "comfy_ui" => Ok(Self::ComfyUi),
@@ -113,6 +119,7 @@ impl InferenceModelType {
       Self::VallEX,
       Self::RerenderAVideo,
       Self::StableDiffusion,
+      Self::ImageGenApi,
       Self::MocapNet,
       Self::StyleTTS2,
       Self::ConvertFbxToGltf,
@@ -140,6 +147,7 @@ mod tests {
       assert_serialization(InferenceModelType::VallEX, "vall_e_x");
       assert_serialization(InferenceModelType::RerenderAVideo, "rerender_a_video");
       assert_serialization(InferenceModelType::StableDiffusion, "stable_diffusion");
+      assert_serialization(InferenceModelType::ImageGenApi, "image_gen_api");
       assert_serialization(InferenceModelType::MocapNet, "mocap_net");
       assert_serialization(InferenceModelType::ComfyUi, "comfy_ui");
       assert_serialization(InferenceModelType::StyleTTS2, "styletts2");
@@ -159,6 +167,7 @@ mod tests {
       assert_eq!(InferenceModelType::VallEX.to_str(), "vall_e_x");
       assert_eq!(InferenceModelType::RerenderAVideo.to_str(), "rerender_a_video");
       assert_eq!(InferenceModelType::StableDiffusion.to_str(), "stable_diffusion");
+      assert_eq!(InferenceModelType::ImageGenApi.to_str(), "image_gen_api");
       assert_eq!(InferenceModelType::MocapNet.to_str(), "mocap_net");
       assert_eq!(InferenceModelType::StyleTTS2.to_str(), "styletts2");
       assert_eq!(InferenceModelType::ConvertFbxToGltf.to_str(), "convert_fbx_gltf");
@@ -176,6 +185,7 @@ mod tests {
       assert_eq!(InferenceModelType::from_str("vall_e_x").unwrap(), InferenceModelType::VallEX);
       assert_eq!(InferenceModelType::from_str("rerender_a_video").unwrap(), InferenceModelType::RerenderAVideo);
       assert_eq!(InferenceModelType::from_str("stable_diffusion").unwrap(), InferenceModelType::StableDiffusion);
+      assert_eq!(InferenceModelType::from_str("image_gen_api").unwrap(), InferenceModelType::ImageGenApi);
       assert_eq!(InferenceModelType::from_str("mocap_net").unwrap(), InferenceModelType::MocapNet);
       assert_eq!(InferenceModelType::from_str("styletts2").unwrap(), InferenceModelType::StyleTTS2);
       assert_eq!(InferenceModelType::from_str("comfy_ui").unwrap(), InferenceModelType::ComfyUi);
@@ -187,17 +197,18 @@ mod tests {
     fn all_variants() {
       // Static check
       let mut variants = InferenceModelType::all_variants();
-      assert_eq!(variants.len(), 13);
+      assert_eq!(variants.len(), 15);
       assert_eq!(variants.pop_first(), Some(InferenceModelType::ComfyUi));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::RvcV2));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::SadTalker));
-      assert_eq!(variants.pop_first(), Some(InferenceModelType::SeedVc));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::SoVitsSvc));
+      assert_eq!(variants.pop_first(), Some(InferenceModelType::SeedVc));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::Tacotron2));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::Vits));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::VallEX));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::RerenderAVideo));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::StableDiffusion));
+      assert_eq!(variants.pop_first(), Some(InferenceModelType::ImageGenApi));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::MocapNet));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::StyleTTS2));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::ConvertFbxToGltf));

crates/_types/enums_public/src/by_table/generic_inference_jobs/public_inference_job_type.rs

@@ -91,6 +91,10 @@ pub enum PublicInferenceJobType {
   /// Re-render a video is a video style transfer algorithm. We developed code
   /// around it, but chose to develop AnimateDiff / ComfyUI support instead.
   RerenderAVideo,
+
+  /// Sora GPT 4o image gen
+  #[serde(rename = "image_gen_api")]
+  ImageGenApi,
 }
 
 /// NB: Legacy API for older code.
@@ -119,6 +123,7 @@ impl PublicInferenceJobType {
       InferenceJobType::Unknown => Self::Unknown,
       InferenceJobType::BevyToWorkflow => Self::BevyToWorkflow,
       InferenceJobType::RerenderAVideo => Self::RerenderAVideo,
+      InferenceJobType::ImageGenApi => Self::ImageGenApi,
     }
   }
   
@@ -140,6 +145,7 @@ impl PublicInferenceJobType {
       Self::SeedVc => InferenceJobType::SeedVc,
       Self::SoVitsSvc => InferenceJobType::SoVitsSvc,
       Self::StableDiffusion => InferenceJobType::StableDiffusion,
+      Self::ImageGenApi => InferenceJobType::ImageGenApi,
       Self::StyleTTS2 => InferenceJobType::StyleTTS2,
       Self::Tacotron2 => InferenceJobType::Tacotron2,
       Self::Unknown => InferenceJobType::Unknown,

crates/desktop/ml_models/Cargo.toml

@@ -0,0 +1,80 @@
+[package]
+name = "ml_models"
+edition = "2021"
+version = "0.0.1"
+authors = ["storyteller"]
+publish = false
+description = ""
+license = ""
+repository = ""
+
+[lib]
+name = "ml_models"
+path = "src/lib.rs"
+crate-type = ["staticlib", "cdylib", "rlib"]
+
+[package.metadata.cargo-udeps.ignore]
+normal = ["my-workspace-hack"]
+
+[dependencies]
+
+# Internal
+filesys = { path = "../../lib/files/filesys" }
+image_algorithms = { path = "../../lib/image_algorithms" }
+memory_caching = { path = "../../lib/caching/memory_caching" }
+ml_weights_registry = { path = "../ml_weights_registry" }
+primitives = { path = "../../lib/data/primitives" }
+
+# Bypass cudarc issues on Windows https://github.com/coreylowman/cudarc/issues/332
+cudarc = "0.13.7"
+
+# External
+anyhow = "1.0"
+base64 = "0.22.1"
+bytes = "1.10"
+candle-core = { path = "../../vendor/candle/candle-core", features = ["cuda"] }
+candle-nn = { path = "../../vendor/candle/candle-nn", features = ["cuda"] }
+candle-transformers = { path = "../../vendor/candle/candle-transformers", features = ["cuda"] }
+const_format = "0.2.34"
+directories = "6.0.0"
+dirs = "6.0.0"
+futures = "0.3"
+half = { version = "2.3.1", features = ["num-traits", "use-intrinsics", "rand_distr"] }
+hf-hub = "0.3.2"
+image = "0.25.5"
+log = "0.4"
+ndarray = "0.15.6"
+nvidia-checker = "0.1.2"
+once_cell = "1.20.3"
+ort = { version = "1.16.2", features = ["cuda"] }
+rand = "0.8.5"
+# NB(bt,2025-03-10): default-features = false on Windows while I debug openssl / rustls
+reqwest = { version = "0.12.12", features = ["stream"], default-features = false }
+serde = { version = "1.0", features = ["derive"] }
+serde_json = "1.0"
+serde_yaml = "0.9"
+tauri = { version = "2.2.4", features = [] }
+tauri-plugin-log = "2"
+tempdir = "0.3.7"
+tempfile = "3.18.0"
+tokenizers = "0.15.2"
+tokio = { version = "1.42", features = ["rt", "rt-multi-thread", "fs"] }
+tokio-util = { version = "0.7", features = ["codec"] }
+
+# tracing-subscriber = { version = "0.3", features = [ "env-filter", "fmt" ] }
+
+[features]
+default = ["cuda"]
+cuda = [
+    "candle-core/cuda",
+    "candle-nn/cuda",
+    "candle-transformers/cuda",
+    # NB(bt,2025-03-10): Building without Flash attention on Windows for now to bypass linker issues
+    "candle-transformers/flash-attn",
+    "candle-core/cudnn"
+]
+
+[dev-dependencies]
+libloading = "0.8.6"
+serial_test = "3.2.0"
+assertor = "0.0.3"

crates/desktop/ml_models/src/lib.rs

@@ -0,0 +1 @@
+pub mod ml;

crates/desktop/ml_models/src/ml/background_removal/mod.rs

@@ -0,0 +1,2 @@
+pub mod onnx_session;
+pub mod remove_image_background;

crates/desktop/ml_models/src/ml/background_removal/onnx_session.rs

@@ -0,0 +1,19 @@
+//! Adapted from https://github.com/dnanhkhoa/rust-background-removal/blob/main/src/onnx.rs
+
+use std::path::Path;
+use ort::{Environment, ExecutionProvider, GraphOptimizationLevel, SessionBuilder};
+
+pub fn onnx_session(onnx_model_file: &Path) -> Result<ort::Session, anyhow::Error> {
+  let environment = Environment::default().into_arc();
+  let session = SessionBuilder::new(&environment)?
+    .with_optimization_level(GraphOptimizationLevel::Level1)? // Configure model optimization level
+    .with_intra_threads(1)? // Configure the number of threads used for inference
+    .with_execution_providers([
+      // Configure execution providers (e.g., CUDA, CoreML, CPU)
+      ExecutionProvider::CUDA(Default::default()),
+      ExecutionProvider::CoreML(Default::default()),
+      ExecutionProvider::CPU(Default::default()),
+    ])?
+    .with_model_from_file(onnx_model_file)?;
+  Ok(session)
+}

crates/desktop/ml_models/src/ml/background_removal/remove_image_background.rs

@@ -0,0 +1,66 @@
+use crate::ml::background_removal::onnx_session::onnx_session;
+use image::{imageops, DynamicImage};
+use ndarray::{Array, CowArray};
+use ort::Value;
+use std::path::Path;
+use image_algorithms::remove_image_border::remove_image_border;
+
+pub async fn remove_image_background<P: AsRef<Path>>(model_path: P, image: DynamicImage) -> anyhow::Result<DynamicImage> {
+  // NB: Real time needed cuDNN 9, but this uses cuDNN 8:
+  // sudo apt-get install libcudnn8-dev libcudnn8
+  // update-alternatives: warning: forcing reinstallation of alternative /usr/include/x86_64-linux-gnu/cudnn_v9.h because link group libcudnn is broken
+  // update-alternatives: using /usr/include/x86_64-linux-gnu/cudnn_v8.h to provide /usr/include/cudnn.h (libcudnn) in manual mode
+  let session = onnx_session(model_path.as_ref())?;
+  let image = process_dynamic_image(&session, image)?;
+  
+  // Crop transparent border cruft away.
+  let image = remove_image_border(&image)?;
+  
+  Ok(image)
+}
+
+fn process_dynamic_image(
+  session: &ort::Session,
+  dynamic_img: DynamicImage,
+) -> Result<DynamicImage, anyhow::Error> {
+  let input_shape = session.inputs[0]
+    .dimensions()
+    .map(|dim| dim.unwrap())
+    .collect::<Vec<usize>>();
+  let input_img = dynamic_img.into_rgba8();
+  let scaling_factor = f32::min(
+    1., // Avoid upscaling
+    f32::min(
+      input_shape[3] as f32 / input_img.width() as f32, // Width ratio
+      input_shape[2] as f32 / input_img.height() as f32, // Height ratio
+    ),
+  );
+  let mut resized_img = imageops::resize(
+    &input_img,
+    input_shape[3] as u32,
+    input_shape[2] as u32,
+    imageops::FilterType::Triangle,
+  );
+  let input_tensor = CowArray::from(
+    Array::from_shape_fn(input_shape, |indices| {
+      let mean = 128.;
+      let std = 256.;
+
+      (resized_img[(indices[3] as u32, indices[2] as u32)][indices[1]] as f32 - mean) / std
+    })
+      .into_dyn(),
+  );
+  let inputs = vec![Value::from_array(session.allocator(), &input_tensor)?];
+  let outputs = session.run(inputs)?;
+  let output_tensor = outputs[0].try_extract::<f32>()?;
+  for (indices, alpha) in output_tensor.view().indexed_iter() {
+    resized_img[(indices[3] as u32, indices[2] as u32)][3] = (alpha * 255.) as u8;
+  }
+  let output_img = imageops::resize(
+    &resized_img,
+    (input_img.width() as f32 * scaling_factor) as u32,
+    (input_img.height() as f32 * scaling_factor) as u32,
+    imageops::FilterType::Triangle,
+  );
+  Ok(DynamicImage::ImageRgba8(output_img))
+}

crates/desktop/ml_models/src/ml/flux/flux_pipeline.rs

@@ -0,0 +1,297 @@
+use std::path::Path;
+use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
+
+use crate::ml::image::dynamic_image_to_tensor::dynamic_image_to_tensor;
+use crate::ml::image::tensor_to_image_buffer::{tensor_to_image_buffer, RgbImage};
+use crate::ml::model_cache::ModelCache;
+use crate::ml::prompt_cache::PromptCache;
+use candle_nn::VarBuilder;
+use crate::ml::model_config::ModelConfig;
+use anyhow::{anyhow, Error as E, Result};
+use candle_core::{DType, IndexOp, Tensor, D, Device, Module};
+use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
+use candle_transformers::models::stable_diffusion::lcm::LCMScheduler;
+use image::DynamicImage;
+use log::{debug, info};
+use rand::Rng;
+use tauri::{AppHandle, Emitter};
+use candle_transformers::models::{clip, t5, flux};
+use tokenizers::Tokenizer;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum FluxModel {
+  Schnell,
+  Dev,
+}
+
+pub struct FluxPipelineArgs<'a> {
+  pub image: &'a DynamicImage,
+  pub prompt: String,
+  pub model_cache: &'a ModelCache,
+  pub prompt_cache: &'a PromptCache,
+  pub app: &'a AppHandle,
+  pub flux_model: FluxModel,
+  pub active_blocks: Option<usize>,
+  pub use_cuda_stream: Option<bool>,
+  pub prefetch_next_batch: Option<bool>,
+  pub use_full_gpu: Option<bool>,
+  pub img2img_strength: Option<f64>,
+  pub seed: Option<u64>,
+  pub use_quantized_model: Option<bool>,
+  pub model_config: &'a ModelConfig,
+  pub flux_dev_or_schnell_path: &'a Path,
+  pub flux_schnell_quantized_gguf: &'a Path,
+  pub flux_autoencoder_path: &'a Path,
+  pub google_t5_model_path: &'a Path,
+  pub google_t5_config_path: &'a Path,
+  pub lmz_t5_tokenizer_path: &'a Path,
+  pub openai_clip_model_path: &'a Path,
+  pub openai_clip_tokenizer_json_path: &'a Path,
+}
+
+pub fn flux_pipeline(args: FluxPipelineArgs<'_>) -> Result<RgbImage> {
+  let FluxPipelineArgs { 
+    prompt, 
+    model_cache, 
+    prompt_cache, 
+    app, 
+    image, 
+    flux_model, 
+    active_blocks, 
+    use_cuda_stream, 
+    prefetch_next_batch, 
+    use_full_gpu, 
+    img2img_strength, 
+    seed, 
+    use_quantized_model, 
+    model_config ,
+    flux_dev_or_schnell_path,
+    flux_schnell_quantized_gguf,
+    flux_autoencoder_path,
+    google_t5_model_path,
+    google_t5_config_path,
+    lmz_t5_tokenizer_path,
+    openai_clip_model_path,
+    openai_clip_tokenizer_json_path,
+  } = args;
+
+  if use_quantized_model.unwrap_or(false) {
+    #[cfg(feature = "cuda")]
+    candle_core::quantized::cuda::set_force_dmmv(false);
+  }
+
+  let height = 1024;
+  let width = 1024;
+
+  let quantized = use_quantized_model.unwrap_or(false);
+  let device = model_config.device.clone();
+  let dtype = device.bf16_default_to_f32();
+
+  info!("Starting image generation with FLUX model:");
+  info!("  Model: {:?}", flux_model);
+  info!("  Prompt: {}", prompt);
+  info!("  Dimensions: {}x{}", width, height);
+  info!("  Device: {:?}", model_config.device);
+
+  let seed_value = seed.unwrap_or_else(|| rand::thread_rng().gen());
+  device.set_seed(seed_value)?;
+  
+  info!("Using seed: {}", seed_value);
+
+  // Setup FLUX model configuration
+  let flux_config = match flux_model {
+    FluxModel::Dev => flux::model::Config::dev(),
+    FluxModel::Schnell => flux::model::Config::schnell(),
+  };
+
+  // Get T5 embeddings
+  let t5_emb = {
+    info!("Generating T5 embeddings for prompt...");
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[google_t5_model_path], dtype , &device)? };
+    let config = std::fs::read_to_string(google_t5_config_path)?;
+    let config: t5::Config = serde_json::from_str(&config)?;
+
+    let mut model = t5::T5EncoderModel::load(vb, &config)?;
+    let tokenizer = Tokenizer::from_file(lmz_t5_tokenizer_path).map_err(E::msg)?;
+
+    let mut tokens = tokenizer.encode(prompt.as_str(), true).map_err(E::msg)?.get_ids().to_vec();
+    tokens.resize(256, 0);
+    let input_token_ids = Tensor::new(&tokens[..], &device)?.unsqueeze(0)?;
+
+    model.forward(&input_token_ids)?
+  };
+  
+  info!("T5 embeddings generated");
+
+  // Get CLIP embeddings
+  let clip_emb = {
+    info!("Generating CLIP embeddings for prompt...");
+    let vb =
+        unsafe { VarBuilder::from_mmaped_safetensors(&[openai_clip_model_path], dtype, &device)? };
+    // https://huggingface.co/openai/clip-vit-large-patch14/blob/main/config.json
+    let config = clip::text_model::ClipTextConfig {
+        vocab_size: 49408,
+        projection_dim: 768,
+        activation: clip::text_model::Activation::QuickGelu,
+        intermediate_size: 3072,
+        embed_dim: 768,
+        max_position_embeddings: 77,
+        pad_with: None,
+        num_hidden_layers: 12,
+        num_attention_heads: 12,
+    };
+    let model =
+        clip::text_model::ClipTextTransformer::new(vb.pp("text_model"), &config)?;
+    let tokenizer = Tokenizer::from_file(openai_clip_tokenizer_json_path).map_err(E::msg)?;
+    let tokens = tokenizer
+        .encode(prompt.as_str(), true)
+        .map_err(E::msg)?
+        .get_ids()
+        .to_vec();
+    let input_token_ids = Tensor::new(&tokens[..], &device)?.unsqueeze(0)?;
+    model.forward(&input_token_ids)?
+  };
+  info!("CLIP embeddings generated");
+
+  // Process input image to tensor
+  info!("Converting input image to tensor...");
+  let img_tensor = dynamic_image_to_tensor(image, &device, dtype)?;
+  info!("Image tensor shape: {:?}", img_tensor.shape());
+
+  // Create initial noise - this is important to match the expected format
+  info!("Generating initial noise for FLUX...");
+  let initial_tensor = flux::sampling::get_noise(1, height, width, &device)?.to_dtype(dtype)?;
+  info!("Initial tensor shape: {:?}", initial_tensor.shape());
+  
+  // Blend initial noise with input image if img2img_strength is specified
+  let initial_tensor = if let Some(strength) = img2img_strength {
+    if strength > 0.0 && strength < 1.0 {
+      info!("Using img2img with strength: {}", strength);
+      // Reshape the input tensor to match the noise tensor dimensions
+      let adjusted_img = if img_tensor.shape() != initial_tensor.shape() {
+        info!("Reshaping input image to match noise dimensions...");
+        // You may need more complex reshaping logic here depending on your use case
+        let (_, c, h, w) = initial_tensor.dims4()?;
+        img_tensor.upsample_nearest2d(h, w)?
+      } else {
+        img_tensor
+      };
+      
+      // Blend the noise and the image according to strength
+      let strength_tensor = Tensor::new(strength as f32, &device)?;
+      let inverse_strength = Tensor::new(1.0 - strength as f32, &device)?;
+      let weighted_noise = (&initial_tensor * &strength_tensor)?;
+      let weighted_img = (&adjusted_img * &inverse_strength)?;
+      (&weighted_noise + &weighted_img)?
+    } else {
+      initial_tensor
+    }
+  } else {
+    initial_tensor
+  };
+  info!("Final initial tensor shape: {:?}", initial_tensor.shape());
+
+  // Create FLUX state
+  info!("Creating FLUX state...");
+  let state = if quantized {
+    flux::sampling::State::new(
+        &t5_emb.to_dtype(DType::F32)?,
+        &clip_emb.to_dtype(DType::F32)?,
+        &initial_tensor.to_dtype(DType::F32)?,
+    )?
+  } else {
+    flux::sampling::State::new(&t5_emb, &clip_emb, &initial_tensor)?
+  };
+  info!("{state:?}");
+
+  // Set appropriate timesteps based on model
+  let timesteps = match flux_model {
+    FluxModel::Dev => flux::sampling::get_schedule(28, Some((state.img.dim(1)?, 0.5, 1.15))),
+    FluxModel::Schnell => flux::sampling::get_schedule(4, None),
+  };
+
+  info!("Starting FLUX denoising with timesteps: {:?}", timesteps);
+
+  // Actual model loading and inference
+  let img_latent = if quantized {
+    // Quantized model path
+    info!("Loading quantized FLUX model...");
+    let model_file = match flux_model {
+      FluxModel::Schnell => flux_schnell_quantized_gguf,
+      FluxModel::Dev => return Err(anyhow!("Quantized model not available for FLUX.1-dev")),
+    };
+
+    info!("Loading quantized model from {}", model_file.to_str().unwrap());
+
+    let vb = candle_transformers::quantized_var_builder::VarBuilder::from_gguf(model_file, &device)?;
+    let mut model = flux::quantized_model::Flux::new(&flux_config, vb)?;
+    info!("Quantized model loaded on {:?}", device);
+
+    // Run denoising
+    info!("Running denoising...");
+    let img_latent = flux::sampling::denoise(&mut model, &state.img, &state.img_ids, &state.txt, &state.txt_ids, &state.vec, &timesteps, 4.)?.to_dtype(dtype)?;
+    info!("Denoising completed");
+    img_latent
+  } else {
+    // Full precision model
+    info!("Loading FLUX model...");
+
+    let mut model = if use_full_gpu.unwrap_or(false) {
+      // Load entire model on GPU
+      info!("Loading entire model directly to GPU...");
+      let gpu_vb = unsafe { VarBuilder::from_mmaped_safetensors(&[flux_dev_or_schnell_path], dtype, &device)? };
+      let mut model = flux::model::Flux::new(&flux_config, gpu_vb)?;
+      model.set_use_device_management(false);
+      model
+    } else {
+      // Use memory-efficient approach with blocks on CPU
+      info!("Using memory-efficient approach with blocks on CPU...");
+      let cpu_device = candle_core::Device::Cpu;
+      let cpu_vb = unsafe { VarBuilder::from_mmaped_safetensors(&[flux_dev_or_schnell_path], dtype, &cpu_device)? };
+      let mut model = flux::model::Flux::new_with_gpu_core(&flux_config, cpu_vb, &device)?;
+
+      // Set active blocks
+      let active_blocks_count = active_blocks.unwrap_or(1);
+      model.set_max_active_blocks(active_blocks_count);
+      info!("Using {} active blocks on GPU", active_blocks_count);
+
+      // Enable prefetching if requested
+      if prefetch_next_batch.unwrap_or(false) {
+        model.set_prefetch_next_batch(true);
+        info!("Prefetching enabled - will overlap transfers with computation");
+      }
+
+      model
+    };
+
+    // Run denoising
+    flux::sampling::denoise(&mut model, &state.img, &state.img_ids, &state.txt, &state.txt_ids, &state.vec, &timesteps, 4.)?
+  };
+
+  info!("FLUX denoising completed, unpacking latent image...");
+  let img_unpacked = flux::sampling::unpack(&img_latent, height, width)?;
+
+  // Load autoencoder for final decoding
+  info!("Loading autoencoder for final decoding...");
+  
+  let ae_config = match flux_model {
+    FluxModel::Dev => flux::autoencoder::Config::dev(),
+    FluxModel::Schnell => flux::autoencoder::Config::schnell(),
+  };
+
+  let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[flux_autoencoder_path], dtype, &device)? };
+  let ae_model = flux::autoencoder::AutoEncoder::new(&ae_config, vb)?;
+
+  info!("Decoding final image...");
+  let img = ae_model.decode(&img_unpacked)?;
+
+  // Convert tensor to image
+  info!("Converting tensor to image...");
+  let img = ((img.clamp(-1f32, 1f32)? + 1.0)? * 127.5)?.to_dtype(candle_core::DType::U8)?;
+
+  // Convert to RGB image
+  let image = tensor_to_image_buffer(&img.i(0)?)?;
+
+  info!("FLUX pipeline completed successfully");
+  Ok(image)
+}

crates/desktop/ml_models/src/ml/flux/mod.rs

@@ -0,0 +1 @@
+pub mod flux_pipeline;

crates/desktop/ml_models/src/ml/mod.rs

@@ -0,0 +1,8 @@
+pub mod background_removal;
+pub mod flux;
+pub mod image;
+pub mod model_cache;
+pub mod model_config;
+pub mod model_file;
+pub mod prompt_cache;
+pub mod stable_diffusion;

crates/desktop/ml_models/src/ml/model_cache.rs

@@ -1,15 +1,14 @@
-use crate::ml::models::unet_model::UNetModel;
 use anyhow::anyhow;
+use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModel;
 use candle_transformers::models::stable_diffusion::vae::AutoEncoderKL;
 use std::sync::{Arc, RwLock};
-
 // TODO: This data structure is gross. Generalize the locking and loading semantics for reuse.
 
 // Simple registry for now. We can build complex machinery that aids in 
 // VRAM utilization, disk space saving, intelligent scheduling, etc. in the future.
 pub struct ModelCache {
   vae: RwLock<Option<Arc<AutoEncoderKL>>>,
-  unet2: RwLock<Option<Arc<UNetModel>>>,
+  unet2: RwLock<Option<Arc<UNet2DConditionModel>>>,
 }
 
 impl ModelCache {
@@ -39,7 +38,7 @@ impl ModelCache {
     }
   }
 
-  pub fn set_unet(&self, vae: Arc<UNetModel>) -> anyhow::Result<()> {
+  pub fn set_unet(&self, vae: Arc<UNet2DConditionModel>) -> anyhow::Result<()> {
     match self.unet2.write() {
       Err(err) => Err(anyhow!("{}", err)),
       Ok(mut lock) => {
@@ -49,7 +48,7 @@ impl ModelCache {
     }
   }
 
-  pub fn get_unet(&self) -> anyhow::Result<Option<Arc<UNetModel>>> {
+  pub fn get_unet(&self) -> anyhow::Result<Option<Arc<UNet2DConditionModel>>> {
     match self.unet2.read() {
       Err(err) => Err(anyhow!("{}", err)),
       Ok(ref lock) => {

crates/desktop/ml_models/src/ml/model_config.rs

@@ -0,0 +1,61 @@
+use crate::ml::model_file::StableDiffusionVersion;
+use candle_core::{DType, Device};
+use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
+use hf_hub::api::sync::Api;
+use log::info;
+
+const DEFAULT_SD_IMAGE_WIDTH: usize = 1024;
+const DEFAULT_SD_IMAGE_HEIGHT: usize = 1024;
+const DEFAULT_SD_NUMERIC_DATATYPE: DType = DType::F32;
+const SD_VERSION: StableDiffusionVersion = StableDiffusionVersion::V1_5;
+
+pub struct ModelConfig {
+
+  /// We only support one device type currently, and we can't multiplex
+  /// over several devices yet.
+  pub device: Device,
+
+  /// The core numeric type to use for models.
+  pub dtype: DType,
+
+  pub sd_version: StableDiffusionVersion,
+  pub sd_config: StableDiffusionConfig,
+
+  /// Probably shouldn't be here.
+  pub hf_api: Api,
+}
+
+impl ModelConfig {
+  pub fn init() -> anyhow::Result<Self> {
+    let device = Device::new_cuda(0)
+      .unwrap_or_else(|e| {
+          info!("CUDA not available ({}), falling back to CPU", e);
+          Device::Cpu
+        });
+
+    let sd_config = match SD_VERSION {
+      StableDiffusionVersion::Turbo => {
+        StableDiffusionConfig::sdxl_turbo(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
+      }
+      StableDiffusionVersion::V1_5 | StableDiffusionVersion::V1_5Inpaint => {
+        StableDiffusionConfig::v1_5(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
+      }
+      StableDiffusionVersion::V2_1 | StableDiffusionVersion::V2Inpaint => {
+        StableDiffusionConfig::v2_1(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
+      }
+      StableDiffusionVersion::Xl | StableDiffusionVersion::XlInpaint => {
+        StableDiffusionConfig::sdxl(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
+      }
+    };
+
+    let hf_api = Api::new()?;
+
+    Ok(Self {
+      device,
+      dtype: DEFAULT_SD_NUMERIC_DATATYPE,
+      sd_version: SD_VERSION,
+      sd_config,
+      hf_api,
+    })
+  }
+}

crates/desktop/ml_models/src/ml/model_file.rs

@@ -0,0 +1,180 @@
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+use candle_transformers::models::stable_diffusion;
+use anyhow::{Error as E, Result};
+use hf_hub::api::sync::Api;
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum StableDiffusionVersion {
+    V1_5,
+    V1_5Inpaint,
+    V2_1,
+    V2Inpaint,
+    Xl,
+    XlInpaint,
+    Turbo,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum ModelFile {
+    Tokenizer,
+    Tokenizer2,
+    Clip,
+    Clip2,
+    Unet,
+    UnetLcm,
+    Vae,
+}
+
+impl StableDiffusionVersion {
+    pub fn repo(&self) -> &'static str {
+        match self {
+            Self::XlInpaint => "diffusers/stable-diffusion-xl-1.0-inpainting-0.1",
+            Self::Xl => "stabilityai/stable-diffusion-xl-base-1.0",
+            Self::V2Inpaint => "stabilityai/stable-diffusion-2-inpainting",
+            Self::V2_1 => "stabilityai/stable-diffusion-2-1",
+            Self::V1_5 => "Lykon/dreamshaper-7",
+            Self::V1_5Inpaint => "stable-diffusion-v1-5/stable-diffusion-inpainting",
+            Self::Turbo => "Lykon/dreamshaper-7",
+        }
+    }
+
+    fn unet_file(&self, use_f16: bool) -> &'static str {
+        match self {
+            Self::V1_5
+            | Self::V1_5Inpaint
+            | Self::V2_1
+            | Self::V2Inpaint
+            | Self::Xl
+            | Self::XlInpaint
+            | Self::Turbo => {
+                if use_f16 {
+                    "unet/diffusion_pytorch_model.fp16.safetensors"
+                } else {
+                    "unet/diffusion_pytorch_model.safetensors"
+                }
+            }
+        }
+    }
+
+    fn vae_file(&self, use_f16: bool) -> &'static str {
+        match self {
+            Self::V1_5
+            | Self::V1_5Inpaint
+            | Self::V2_1
+            | Self::V2Inpaint
+            | Self::Xl
+            | Self::XlInpaint
+            | Self::Turbo => {
+                if use_f16 {
+                    "vae/diffusion_pytorch_model.fp16.safetensors"
+                } else {
+                    "vae/diffusion_pytorch_model.safetensors"
+                }
+            }
+        }
+    }
+
+    fn clip_file(&self, use_f16: bool) -> &'static str {
+        match self {
+            Self::V1_5
+            | Self::V1_5Inpaint
+            | Self::V2_1
+            | Self::V2Inpaint
+            | Self::Xl
+            | Self::XlInpaint
+            | Self::Turbo => {
+                if use_f16 {
+                    "text_encoder/model.fp16.safetensors"
+                } else {
+                    "text_encoder/model.safetensors"
+                }
+            }
+        }
+    }
+
+    fn clip2_file(&self, use_f16: bool) -> &'static str {
+        match self {
+            Self::V1_5
+            | Self::V1_5Inpaint
+            | Self::V2_1
+            | Self::V2Inpaint
+            | Self::Xl
+            | Self::XlInpaint
+            | Self::Turbo => {
+                if use_f16 {
+                    "text_encoder_2/model.fp16.safetensors"
+                } else {
+                    "text_encoder_2/model.safetensors"
+                }
+            }
+        }
+    }
+}
+
+impl ModelFile {
+    pub fn get(
+        &self,
+        filename: Option<String>,
+        version: StableDiffusionVersion,
+        use_f16: bool,
+    ) -> Result<std::path::PathBuf> {
+        match filename {
+            Some(filename) => Ok(std::path::PathBuf::from(filename)),
+            None => {
+                let (repo, path) = match self {
+                    Self::Tokenizer => {
+                        let tokenizer_repo = match version {
+                            StableDiffusionVersion::V1_5
+                            | StableDiffusionVersion::V2_1
+                            | StableDiffusionVersion::V1_5Inpaint
+                            | StableDiffusionVersion::V2Inpaint => "openai/clip-vit-base-patch32",
+                            StableDiffusionVersion::Xl
+                            | StableDiffusionVersion::XlInpaint
+                            | StableDiffusionVersion::Turbo => {
+                                // This seems similar to the patch32 version except some very small
+                                // difference in the split regex.
+                                "openai/clip-vit-large-patch14"
+                            }
+                        };
+                        (tokenizer_repo, "tokenizer.json")
+                    }
+                    Self::Tokenizer2 => {
+                        ("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", "tokenizer.json")
+                    }
+                    Self::Clip => (version.repo(), version.clip_file(use_f16)),
+                    Self::Clip2 => (version.repo(), version.clip2_file(use_f16)),
+                    Self::UnetLcm => (
+                        "SimianLuo/LCM_Dreamshaper_v7",
+                        if use_f16 {
+                            "unet/diffusion_pytorch_model.fp16.safetensors"
+                        } else {
+                            "unet/diffusion_pytorch_model.safetensors"
+                        },
+                    ),
+                    Self::Unet => (version.repo(), version.unet_file(use_f16)),
+                    Self::Vae => {
+                        // Override for SDXL when using f16 weights.
+                        // See https://github.com/huggingface/candle/issues/1060
+                        if matches!(
+                            version,
+                            StableDiffusionVersion::Xl | StableDiffusionVersion::Turbo,
+                        ) && use_f16
+                        {
+                            (
+                                "madebyollin/sdxl-vae-fp16-fix",
+                                "diffusion_pytorch_model.safetensors",
+                            )
+                        } else {
+                            (version.repo(), version.vae_file(use_f16))
+                        }
+                    }
+                };
+                let filename = Api::new()?.model(repo.to_string()).get(path)?;
+                Ok(filename)
+            }
+        }
+    }
+}

crates/desktop/ml_models/src/ml/stable_diffusion/infer_clip_text_embeddings.rs

@@ -0,0 +1,158 @@
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+use candle_transformers::models::stable_diffusion;
+use std::path::{Path, PathBuf};
+
+use crate::ml::model_file::{ModelFile, StableDiffusionVersion};
+use ml_weights_registry::weights_registry::weights::{CLIP_JSON, SDXL_TURBO_CLIP_TEXT_ENCODER, SDXL_TURBO_CLIP_TEXT_ENCODER_2};
+use anyhow::Error as E;
+use candle_core::{DType, Device, Module, Tensor, D};
+use log::info;
+use tokenizers::Tokenizer;
+
+
+pub struct ClipArgs<'a, P: AsRef<Path>,  Q: AsRef<Path>> {
+  pub prompt: &'a str, 
+  pub uncond_prompt: &'a str, 
+  pub tokenizer: Option<String>, 
+  pub clip_weights: Option<String>, 
+  pub clip2_weights: Option<String>, 
+  pub sd_version: StableDiffusionVersion, 
+  pub sd_config: &'a stable_diffusion::StableDiffusionConfig, 
+  pub use_f16: bool, 
+  pub device: &'a Device, 
+  pub dtype: DType, 
+  pub use_guide_scale: bool, 
+  pub clip_json_path: P,
+  pub clip_weights_path: Q,
+}
+
+#[allow(clippy::too_many_arguments)]
+pub fn infer_clip_text_embeddings<P: AsRef<Path>, Q: AsRef<Path>>(args: ClipArgs<'_, P, Q>) -> anyhow::Result<Tensor> {
+  let ClipArgs {
+    prompt, 
+    uncond_prompt, 
+    tokenizer, 
+    clip_weights, 
+    clip2_weights, 
+    sd_version, 
+    sd_config, 
+    use_f16, 
+    device, 
+    dtype, 
+    use_guide_scale, 
+    clip_json_path,
+    clip_weights_path,
+  } = args;
+
+  info!("Preparing text embeddings... for {:?}", sd_version);
+
+  let which = match sd_version {
+    StableDiffusionVersion::Xl | StableDiffusionVersion::XlInpaint | StableDiffusionVersion::Turbo => vec![true, false],
+    _ => vec![true],
+  };
+
+  let text_embeddings = which
+    .iter()
+    .map(|first| {
+      do_infer_clip_text_embeddings(
+        &prompt,
+        &uncond_prompt,
+        None, // tokenizer
+        clip_weights.clone(),
+        None, // clip2_weights
+        sd_version,
+        &sd_config,
+        false, // use_f16
+        &device,
+        dtype,
+        use_guide_scale,
+        *first,
+        clip_json_path.as_ref(),
+        clip_weights_path.as_ref(),
+      )
+    })
+    .collect::<anyhow::Result<Vec<_>>>()?;
+
+  let text_embeddings = Tensor::cat(&text_embeddings, D::Minus1)?;
+
+  Ok(text_embeddings)
+}
+
+#[allow(clippy::too_many_arguments)]
+fn do_infer_clip_text_embeddings(
+  prompt: &str, 
+  uncond_prompt: &str, 
+  tokenizer: Option<String>, 
+  clip_weights: Option<String>, 
+  clip2_weights: Option<String>, 
+  sd_version: StableDiffusionVersion, 
+  sd_config: &stable_diffusion::StableDiffusionConfig, 
+  use_f16: bool, 
+  device: &Device, 
+  dtype: DType, 
+  use_guide_scale: bool, 
+  first: bool, 
+  clip_json_path: &Path,
+  clip_weights_path: &Path,
+) -> anyhow::Result<Tensor> {
+  info!("do_infer_clip_text_embeddings called with args {:?}", (prompt, uncond_prompt, tokenizer.clone().unwrap_or_else(|| "None".to_string()), clip_weights.clone().unwrap_or_else(|| "None".to_string()), clip2_weights.clone().unwrap_or_else(|| "None".to_string()), sd_version, sd_config, use_f16, device, dtype, use_guide_scale, first));
+
+  let tokenizer = clip_json_path;
+
+  info!("Loading Clip Tokenizer path: {:?}", tokenizer);
+
+  let tokenizer = Tokenizer::from_file(tokenizer).map_err(E::msg)?;
+
+  let pad_id = match &sd_config.clip.pad_with {
+    Some(padding) => *tokenizer.get_vocab(true).get(padding.as_str()).unwrap(),
+    None => *tokenizer.get_vocab(true).get("<|endoftext|>").unwrap(),
+  };
+
+  println!("Running with prompt \"{prompt}\".");
+  let mut tokens = tokenizer.encode(prompt, true).map_err(E::msg)?.get_ids().to_vec();
+  if tokens.len() > sd_config.clip.max_position_embeddings {
+    anyhow::bail!("the prompt is too long, {} > max-tokens ({})", tokens.len(), sd_config.clip.max_position_embeddings)
+  }
+  while tokens.len() < sd_config.clip.max_position_embeddings {
+    tokens.push(pad_id)
+  }
+  let tokens = Tensor::new(tokens.as_slice(), device)?.unsqueeze(0)?;
+
+  println!("Building the Clip transformer.");
+  
+  // TODO(bt,2025-03-23): These used to be two different clip models. We should only need one.
+  let clip_weights = if first {
+    println!(">>>>> CLIP 1");
+    clip_weights_path
+  } else {
+    println!(">>>>> CLIP 2");
+    clip_weights_path
+  };
+  
+  let clip_config = if first { &sd_config.clip } else { sd_config.clip2.as_ref().unwrap() };
+  let text_model = stable_diffusion::build_clip_transformer(clip_config, clip_weights, device, DType::F32)?;
+
+  let text_embeddings = text_model.forward(&tokens)?;
+
+  let text_embeddings = if use_guide_scale {
+    let mut uncond_tokens = tokenizer.encode(uncond_prompt, true).map_err(E::msg)?.get_ids().to_vec();
+    if uncond_tokens.len() > sd_config.clip.max_position_embeddings {
+      anyhow::bail!("the negative prompt is too long, {} > max-tokens ({})", uncond_tokens.len(), sd_config.clip.max_position_embeddings)
+    }
+    while uncond_tokens.len() < sd_config.clip.max_position_embeddings {
+      uncond_tokens.push(pad_id)
+    }
+
+    let uncond_tokens = Tensor::new(uncond_tokens.as_slice(), device)?.unsqueeze(0)?;
+    let uncond_embeddings = text_model.forward(&uncond_tokens)?;
+
+    Tensor::cat(&[uncond_embeddings, text_embeddings], 0)?.to_dtype(dtype)?
+  } else {
+    text_embeddings.to_dtype(dtype)?
+  };
+  Ok(text_embeddings)
+}

crates/desktop/ml_models/src/ml/stable_diffusion/lcm_pipeline.rs

@@ -0,0 +1,281 @@
+use std::path::{Path, PathBuf};
+use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
+
+use anyhow::{anyhow, Error as E, Result};
+use candle_core::{DType, Device, IndexOp, Tensor, D};
+use candle_nn::VarBuilder;
+use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
+use candle_transformers::models::stable_diffusion::lcm::{LCMScheduler, LCMSchedulerConfig};
+use candle_transformers::models::stable_diffusion::unet_2d::BlockConfig;
+use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModel;
+use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModelConfig;
+use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
+use crate::ml::image::dynamic_image_to_tensor::dynamic_image_to_tensor;
+use crate::ml::image::tensor_to_image_buffer::{tensor_to_image_buffer, RgbImage};
+use crate::ml::model_cache::ModelCache;
+use crate::ml::model_config::ModelConfig;
+use crate::ml::model_file::{ModelFile, StableDiffusionVersion};
+use crate::ml::prompt_cache::PromptCache;
+use crate::ml::stable_diffusion::get_vae_scale::get_vae_scale;
+use crate::ml::stable_diffusion::infer_clip_text_embeddings::{infer_clip_text_embeddings, ClipArgs};
+use crate::ml::stable_diffusion::remap_lcm_strength_range::remap_lcm_strength_range;
+use image::DynamicImage;
+use log::{debug, error, info};
+use ml_weights_registry::weights_registry::weights::{LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16, LYKON_DEAMSHAPER_7_VAE, SIMIANLUO_LCM_DREAMSHAPER_V7_UNET};
+use rand::Rng;
+
+const DEFAULT_STRENGTH : f64 = 75.0;
+
+const DEFAULT_SEED : u64 = 42;
+
+pub struct Args<'a, P1: AsRef<Path>, P2: AsRef<Path>, P3: AsRef<Path>, P4: AsRef<Path>> {
+  pub image: &'a DynamicImage,
+  pub prompt: String,
+  pub uncond_prompt: String,
+  pub cfg_scale: Option<f64>,
+  pub img2img_strength: Option<f64>,
+  pub model_cache: &'a ModelCache,
+  pub prompt_cache: &'a PromptCache,
+  pub use_flash_attn: bool,
+  pub model_config: &'a ModelConfig,
+  pub maybe_seed: Option<u64>,
+  pub scheduler_steps: usize,
+  pub vae_path: P1,
+  pub unet_path: P2,
+  pub clip_json_path: P3,
+  pub clip_weights_path: P4,
+}
+
+pub fn lcm_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>) -> Result<RgbImage> 
+  where P1: AsRef<Path>, P2: AsRef<Path>, P3: AsRef<Path>, P4: AsRef<Path>
+{
+  let Args { 
+    prompt, 
+    uncond_prompt, 
+    cfg_scale,
+    img2img_strength, 
+    model_cache, 
+    prompt_cache, 
+    image, 
+    use_flash_attn ,
+    model_config: ModelConfig {
+      device,
+      dtype,
+      sd_version,
+      sd_config,
+      hf_api,
+    },
+    maybe_seed,
+    scheduler_steps,
+    vae_path,
+    unet_path,
+    clip_json_path,
+    clip_weights_path,
+  } = args;
+  
+  let dtype = *dtype;
+  let sd_version = *sd_version;
+
+  let img2img_strength = img2img_strength
+    .map(|s| remap_lcm_strength_range(s))
+    .unwrap_or(DEFAULT_STRENGTH);
+  
+  let img2img_strength = img2img_strength / 100.0;
+
+  let use_f16 = true;
+
+  // Use LCM Scheduler instead of Euler Ancestral for better speed and quality
+  let mut scheduler = LCMScheduler::new(scheduler_steps, img2img_strength, LCMSchedulerConfig::default())?;
+
+  //let seed = maybe_seed.unwrap_or_else(|| rand::thread_rng().gen());
+  let seed = maybe_seed.unwrap_or(DEFAULT_SEED);
+
+  device.set_seed(seed)?;
+
+  let guidance_scale = match cfg_scale {
+    Some(guidance_scale) => guidance_scale,
+    None => match sd_version {
+      StableDiffusionVersion::V1_5 | StableDiffusionVersion::V2_1 | StableDiffusionVersion::Xl => 7.5,
+      StableDiffusionVersion::Turbo => 0.,
+      _ => 0., // NB: Not sure what the other model families should use, so sticking with "0"
+    },
+  };
+
+  // let use_guide_scale = guidance_scale > 1.0;
+  let use_guide_scale = false;
+
+  let maybe_cached = prompt_cache.get_copy(&prompt)?;
+
+  let mut text_embeddings = if let Some(tensor) = maybe_cached {
+    tensor
+  } else {
+    info!("Prompt is NOT cached! Calculating embedding...");
+    let tensor = infer_clip_text_embeddings(ClipArgs {
+      prompt: & prompt,
+      uncond_prompt: &uncond_prompt,
+      tokenizer: None, // tokenizer
+      clip_weights: None,
+      clip2_weights: None, // clip2_weights
+      sd_version,
+      sd_config: &sd_config,
+      use_f16, // use_f16
+      device,
+      dtype,
+      use_guide_scale,
+      clip_json_path,
+      clip_weights_path,
+    })?;
+
+    prompt_cache.store_copy(&prompt, &tensor)?;
+    tensor
+  };
+
+  debug!("Text embeddings shape: {:?}", text_embeddings.shape());
+
+  debug!("Loading input image into tensor...");
+
+  let input_image = dynamic_image_to_tensor(image, &device, dtype)?;
+
+  debug!("Reference image shape: {:?}", input_image.shape());
+
+  let vae_scale = get_vae_scale(sd_version);
+
+  let maybe_vae = model_cache.get_vae()?;
+
+  let vae = match maybe_vae {
+    Some(vae) => vae,
+    None => {
+      info!("Building VAE model from file {:?}...", vae_path.as_ref());
+      
+      let vae = sd_config.build_vae(vae_path, &device, dtype)?;
+
+      let vae = Arc::new(vae);
+
+      model_cache.set_vae(vae.clone())?;
+
+      vae
+    }
+  };
+
+  let maybe_unet = model_cache.get_unet()?;
+
+  let unet = match maybe_unet {
+    Some(unet) => unet,
+    None => {
+      info!("Building UNET model from file {:?}...", unet_path.as_ref());
+
+      let in_channels = match sd_version {
+        StableDiffusionVersion::XlInpaint | StableDiffusionVersion::V2Inpaint | StableDiffusionVersion::V1_5Inpaint => 9,
+        _ => 4,
+      };
+
+      // For LCM models, ensure that:
+      // 1. We're using the correct prediction type (epsilon)
+      // 2. The cross-attention dimension is correct (768 for Dreamshaper)
+      // 3. The in_channels matches the expected input (4 for standard, 9 for inpainting)
+      let unet_config = UNet2DConditionModelConfig {
+        // LCM specific config values
+        cross_attention_dim: 768,      // Dreamshaper v7 uses 768
+        time_cond_proj_dim: Some(256), // LCM models use 256-dimensional guidance embeddings
+        // Rest of configuration...
+        center_input_sample: false,
+        flip_sin_to_cos: true,
+        freq_shift: 0.0,
+        blocks: vec![
+          BlockConfig { out_channels: 320, use_cross_attn: Some(1), attention_head_dim: 8 },
+          BlockConfig { out_channels: 640, use_cross_attn: Some(1), attention_head_dim: 8 },
+          BlockConfig { out_channels: 1280, use_cross_attn: Some(1), attention_head_dim: 8 },
+          BlockConfig { out_channels: 1280, use_cross_attn: None, attention_head_dim: 8 }
+        ],
+        layers_per_block: 2,
+        downsample_padding: 1,
+        mid_block_scale_factor: 1.0,
+        norm_num_groups: 32,
+        norm_eps: 1e-5,
+        sliced_attention_size: None,
+        use_linear_projection: false,
+        time_embed_dim: Some(1280),
+        // time_embed_dim: Some(256),
+      };
+
+      // Load the model directly from the safetensors file
+      let vs_unet = unsafe { VarBuilder::from_mmaped_safetensors(&[unet_path], dtype, &device)? };
+
+      let unet = UNet2DConditionModel::new(vs_unet, in_channels, 4, use_flash_attn, unet_config)?;
+
+      let unet = Arc::new(unet);
+
+      model_cache.set_unet(unet.clone())?;
+
+      unet
+    },
+  };
+
+  let init_latent_dist: DiagonalGaussianDistribution = vae.encode(&input_image)?;
+
+  // Create LCM guidance scale embeddings
+  let embedding_dim = 1280;
+  let guidance_scale_embedding = scheduler.get_guidance_scale_embedding(guidance_scale, embedding_dim, &device, dtype)?;
+
+  // Generate latents from input image using the LCM approach
+  debug!("Generating latents from input image...");
+  let init_latents = (init_latent_dist.sample()? * vae_scale)?;
+
+  debug!("Initial latents shape: {:?}", init_latents.shape());
+
+  // Calculate timesteps for LCM with proper img2img handling
+  let timesteps = LCMScheduler::get_timesteps_for_steps(scheduler_steps, img2img_strength);
+  let t_start = timesteps[0]; // First denoising step
+
+  // Add noise at the appropriate timestep for img2img
+  debug!("Adding noise to latents for t_start={}", t_start);
+  let noise = init_latents.randn_like(0f64, 1f64)?;
+  let latents = scheduler.add_noise(&init_latents, noise, t_start)?;
+
+  let mut latents = latents;
+  let timesteps: Vec<_> = scheduler.timesteps().iter().copied().collect();
+
+  for (timestep_index, &timestep) in timesteps.iter().enumerate() {
+    info!("Processing step {}/{} @ timestamp = {}", timestep_index + 1, timesteps.len(), timestep);
+
+    let latent_model_input = if use_guide_scale { Tensor::cat(&[&latents, &latents], 0)? } else { latents.clone() };
+
+    let latent_model_input = scheduler.scale_model_input(latent_model_input, timestep);
+
+    // Use the guidance scale embedding in the UNet inference with the proper method name
+    let mut noise_pred = match unet.forward_with_guidance(&latent_model_input, timestep as f64, &text_embeddings, Some(&guidance_scale_embedding)) {
+      Ok(pred) => pred,
+      Err(e) => {
+        error!("UNet inference failed with error: {}", e);
+        return Err(anyhow::anyhow!("UNet inference failed: {}", e));
+      },
+    };
+
+    if use_guide_scale {
+      let chunks = noise_pred.chunk(2, 0)?;
+      let (noise_pred_uncond, noise_pred_text) = (&chunks[0], &chunks[1]);
+      noise_pred = (noise_pred_uncond + (guidance_scale * (noise_pred_text - noise_pred_uncond)?)?)?;
+    }
+
+    // Apply the LCM scheduler step
+    latents = scheduler.step(&noise_pred, timestep, &latents, timestep_index, scheduler_steps)?;
+  }
+
+  debug!("Diffusion process completed, decoding image...");
+  let image = vae.decode(&(latents / vae_scale)?)?;
+
+  debug!("Scaling image back");
+  let image = ((image / 2.)? + 0.5)?;
+
+  debug!("Normalized image values");
+  let image = (image.clamp(0f32, 1.)? * 255.)?;
+
+  debug!("Convert to int8");
+  let image = image.to_dtype(DType::U8)?;
+
+  debug!("Converted to 8-bit format");
+
+  let image = tensor_to_image_buffer(&image.i(0)?)?;
+
+  Ok(image)
+}

crates/desktop/ml_models/src/ml/stable_diffusion/mod.rs

@@ -1,4 +1,5 @@
 pub mod get_vae_scale;
 pub mod infer_clip_text_embeddings;
 pub mod stable_diffusion_pipeline;
-pub mod lcm_pipeline;
+pub mod lcm_pipeline;
+mod remap_lcm_strength_range;

crates/desktop/ml_models/src/ml/stable_diffusion/remap_lcm_strength_range.rs

@@ -0,0 +1,54 @@
+use primitives::numerics::remap_range_f64::{remap_range_f64, Args};
+
+/// Our LCM implementation uses strength [13, 100]. Strength below 13 will take 8 steps, 
+/// so we need to limit ourselves to this range from an input range of [0, 100].
+pub fn remap_lcm_strength_range(value: f64) -> f64 {
+  remap_range_f64(Args {
+    old_min: 0.0,
+    old_max: 100.0,
+    new_min: 13.0,
+    new_max: 100.0,
+    value,
+  })
+}
+
+#[cfg(test)]
+mod test {
+  use crate::ml::stable_diffusion::remap_lcm_strength_range::remap_lcm_strength_range;
+  use assertor::{assert_that, FloatAssertion};
+
+  #[test]
+  fn old_max() {
+    assert_eq!(remap_lcm_strength_range(100.0), 100.0);
+  }
+  
+  #[test]
+  fn old_min() {
+    assert_eq!(remap_lcm_strength_range(0.0), 13.0);
+  }
+
+  #[test]
+  fn under_bounds() {
+    assert_eq!(remap_lcm_strength_range(-100.0), 13.0);
+  }
+  
+  #[test]
+  fn over_bounds() {
+    assert_eq!(remap_lcm_strength_range(200.0), 100.0);
+  }
+
+  #[test]
+  fn other_values() {
+    assert_that!(remap_lcm_strength_range(0.0)).with_abs_tol(0.001).is_approx_equal_to(13.000);
+    assert_that!(remap_lcm_strength_range(5.0)).with_abs_tol(0.001).is_approx_equal_to(17.350);
+    assert_that!(remap_lcm_strength_range(11.0)).with_abs_tol(0.001).is_approx_equal_to(22.570);
+    assert_that!(remap_lcm_strength_range(13.0)).with_abs_tol(0.001).is_approx_equal_to(24.310);
+    assert_that!(remap_lcm_strength_range(14.0)).with_abs_tol(0.001).is_approx_equal_to(25.180);
+    assert_that!(remap_lcm_strength_range(20.0)).with_abs_tol(0.001).is_approx_equal_to(30.400);
+    assert_that!(remap_lcm_strength_range(25.0)).with_abs_tol(0.001).is_approx_equal_to(34.750);
+    assert_that!(remap_lcm_strength_range(50.0)).with_abs_tol(0.001).is_approx_equal_to(56.500);
+    assert_that!(remap_lcm_strength_range(75.0)).with_abs_tol(0.001).is_approx_equal_to(78.250);
+    assert_that!(remap_lcm_strength_range(99.0)).with_abs_tol(0.001).is_approx_equal_to(99.130);
+    assert_that!(remap_lcm_strength_range(100.0)).with_abs_tol(0.001).is_approx_equal_to(100.000);
+  }
+}

crates/desktop/ml_models/src/ml/stable_diffusion/stable_diffusion_pipeline.rs

@@ -1,76 +1,78 @@
+use std::path::Path;
 use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
 
-use crate::events::notification_event::{NotificationModelType, NotificationEvent};
 use crate::ml::image::dynamic_image_to_tensor::dynamic_image_to_tensor;
 use crate::ml::image::tensor_to_image_buffer::{tensor_to_image_buffer, RgbImage};
 use crate::ml::model_cache::ModelCache;
 use crate::ml::model_file::StableDiffusionVersion;
-use crate::ml::model_type::ModelType;
-use crate::ml::models::unet_model::UNetModel;
 use crate::ml::prompt_cache::PromptCache;
 use crate::ml::stable_diffusion::get_vae_scale::get_vae_scale;
-use crate::ml::stable_diffusion::infer_clip_text_embeddings::infer_clip_text_embeddings;
-use crate::state::app_config::AppConfig;
+use crate::ml::stable_diffusion::infer_clip_text_embeddings::{infer_clip_text_embeddings, ClipArgs};
 use anyhow::{anyhow, Error as E, Result};
-use candle_core::{DType, IndexOp, Tensor, D};
+use candle_core::{DType, Device, IndexOp, Tensor, D};
 use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
+use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
 use image::DynamicImage;
 use log::info;
 use rand::Rng;
-use tauri::{AppHandle, Emitter};
-use crate::state::app_dir::AppDataRoot;
 
-pub struct Args<'a> {
+pub struct Args<'a, P: AsRef<Path>, Q: AsRef<Path>, R: AsRef<Path>, S: AsRef<Path>> {
     pub image: &'a DynamicImage,
     pub prompt: String,
     pub uncond_prompt: String,
     pub cfg_scale: Option<f64>,
     pub i2i_strength: Option<u8>,
-    pub configs: &'a AppConfig,
     pub model_cache: &'a ModelCache,
     pub prompt_cache: &'a PromptCache,
-    pub app_data_root: &'a AppDataRoot,
-    pub app: &'a AppHandle,
+    pub sd_version: StableDiffusionVersion,
+    pub sd_config: StableDiffusionConfig,
+    pub device: &'a Device,
+    pub dtype: DType,
+    pub maybe_seed: Option<u64>,
+    pub scheduler_steps: usize,
+    pub sdxl_turbo_vae_path: P,
+    pub sdxl_turbo_unet_path: Q,
+    pub clip_json_path: R,
+    pub clip_weights_path: S,
 }
 
-pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
+pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>) -> Result<RgbImage> 
+  where P1: AsRef<Path>, P2: AsRef<Path>, P3: AsRef<Path>, P4: AsRef<Path>
+{
     let Args { 
         prompt, 
         uncond_prompt, 
         cfg_scale, 
         i2i_strength, 
-        configs, 
         model_cache, 
         prompt_cache, 
-        app, 
         image,
-        app_data_root,
+        sd_version,
+        sd_config,
+        device,
+        dtype,
+        maybe_seed,
+        scheduler_steps,
+        sdxl_turbo_vae_path,
+        sdxl_turbo_unet_path,
+        clip_json_path,
+        clip_weights_path,
     } = args;
     
-    let weights_dir = app_data_root.weights_dir();
-    
-    println!("Starting image generation with the following configuration:");
-    println!("  Model: {:?}", configs.sd_version);
-    println!("  Prompt: {}", prompt);
-    println!("  Steps: {}", configs.scheduler_steps);
-    println!("  Device: {:?}", configs.device);
-
-    println!("Model dimensions: {}x{}", configs.sd_config.width, configs.sd_config.height);
-
     // TODO(bt,2025-02-18): The scheduler is `EulerAncestralDiscreteScheduler`, but we may want to port an LCM scheduler.
     //  This is a target for performance improvement
     //  See: https://github.com/huggingface/candle/issues/1331
-    let mut scheduler = configs.sd_config.build_scheduler(
-        configs.scheduler_steps)?;
+    let mut scheduler = sd_config.build_scheduler(scheduler_steps)?;
 
-    let seed = configs.seed.unwrap_or_else(|| rand::thread_rng().gen());
-    configs.device.set_seed(seed)?;
+    let seed = maybe_seed.unwrap_or_else(|| rand::thread_rng().gen());
+    
+    device.set_seed(seed)?;
 
     info!("Using seed: {}", seed);
 
     let guidance_scale = match cfg_scale {
         Some(guidance_scale) => guidance_scale,
-        None => match configs.sd_version {
+        None => match sd_version {
             StableDiffusionVersion::V1_5
             | StableDiffusionVersion::V2_1
             | StableDiffusionVersion::Xl => 7.5,
@@ -91,20 +93,21 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
     } else {
         
         info!("Prompt is NOT cached! Calculating embedding...");
-        let tensor = infer_clip_text_embeddings(
-            &prompt,
-            &uncond_prompt,
-            None, // tokenizer
-            None, // clip_weights
-            None, // clip2_weights
-            configs.sd_version,
-            &configs.sd_config,
-            false, // use_f16
-            &configs.device,
-            configs.dtype,
+        let tensor = infer_clip_text_embeddings(ClipArgs {
+            prompt: &prompt,
+            uncond_prompt: &uncond_prompt,
+            tokenizer: None, // tokenizer
+            clip_weights: None, // clip_weights
+            clip2_weights: None, // clip2_weights
+            sd_version: sd_version,
+            sd_config: &sd_config,
+            use_f16: false, // use_f16
+            device: device,
+            dtype: dtype,
             use_guide_scale,
-            weights_dir,
-        )?;
+            clip_json_path: clip_json_path.as_ref(),
+            clip_weights_path: clip_weights_path.as_ref(),
+        })?;
         prompt_cache.store_copy(&prompt, &tensor)?;
         tensor
     };
@@ -113,14 +116,11 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
 
     println!("Loading input image into tensor...");
 
-    let input_image = dynamic_image_to_tensor(
-        image,
-        &configs.device,
-        configs.dtype)?;
+    let input_image = dynamic_image_to_tensor(image, device, dtype)?;
 
     println!("Reference image shape: {:?}", input_image.shape());
 
-    let vae_scale = get_vae_scale(configs.sd_version);
+    let vae_scale = get_vae_scale(sd_version);
 
     let maybe_vae = model_cache.get_vae()?;
     
@@ -129,33 +129,14 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
         None => {
             info!("No vae found in cache; loading...");
             
-            let mut notify_download_complete = false;
-            //if !vae_file.exists() {
-            if true {
-                notify_download_complete = true;
-                app.emit("notification", NotificationEvent::ModelDownloadStarted {
-                    model_name: "sdxl-turbo-vae",
-                    model_type: NotificationModelType::Vae,
-                })?;
-            }
+            let vae_file = sdxl_turbo_vae_path.as_ref();
             
-            let vae_file = weights_dir.model_path(&ModelType::SdxlTurboVae);
-            
-            let vae = configs
-              .sd_config
-              .build_vae(vae_file, &configs.device, configs.dtype)?;
+            let vae = sd_config.build_vae(vae_file, &device, dtype)?;
 
             let vae = Arc::new(vae);
             
             model_cache.set_vae(vae.clone())?;
-            
-            if notify_download_complete {
-                app.emit("notification", NotificationEvent::ModelDownloadComplete {
-                    model_name: "sdxl-turbo-vae",
-                    model_type: NotificationModelType::Vae,
-                })?;
-            }
-            
+
             vae
         }
     };
@@ -167,32 +148,15 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
         None => {
             info!("No unet found in cache; loading...");
 
-            let mut notify_download_complete = false;
-            //if !unet_file.exists() {
-            if true {
-                notify_download_complete = true;
-                app.emit("notification", NotificationEvent::ModelDownloadStarted {
-                    model_name: "sdxl-turbo-unet",
-                    model_type: NotificationModelType::Unet,
-                })?;
-            }
-            
-            let unet_file = weights_dir.model_path(&ModelType::SdxlTurboUnet);
+            let unet_weights_file = sdxl_turbo_unet_path.as_ref();
 
-            let unet = UNetModel::new(&configs.sd_config, unet_file, &configs.device, configs.dtype)
+            let unet = sd_config.build_unet(unet_weights_file, &device, 4, true, dtype)
               .map_err(|err| anyhow!("error initializing unet model: {:?}", err))?;
-            
+
             let unet = Arc::new(unet);
 
             model_cache.set_unet(unet.clone())?;
 
-            if notify_download_complete {
-                app.emit("notification", NotificationEvent::ModelDownloadComplete {
-                    model_name: "sdxl-turbo-unet",
-                    model_type: NotificationModelType::Unet,
-                })?;
-            }
-
             unet
         }
     };
@@ -216,9 +180,9 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
     };
 
     let t_start = {
-        let start = configs.scheduler_steps - (configs.scheduler_steps as f64 * img2img_strength) as usize;
+        let start = scheduler_steps - (scheduler_steps as f64 * img2img_strength) as usize;
 
-        println!("Starting from step {} of {} (strength: {})", start, configs.scheduler_steps, img2img_strength);
+        println!("Starting from step {} of {} (strength: {})", start, scheduler_steps, img2img_strength);
         start
     };
 
@@ -257,7 +221,7 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
 
         let latent_model_input = scheduler.scale_model_input(latent_model_input, timestep)?;
 
-        let mut noise_pred = match unet.inference(&latent_model_input, timestep as f64, &text_embeddings) {
+        let mut noise_pred = match unet.forward(&latent_model_input, timestep as f64, &text_embeddings) {
             Ok(pred) => pred,
             Err(e) => {
                 println!("UNet inference failed with error: {}", e);

crates/desktop/ml_weights_registry/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "ml_weights_registry"
+edition = "2021"
+version = "0.0.1"
+authors = ["storyteller"]
+publish = false
+description = ""
+license = ""
+repository = ""
+
+[lib]
+name = "ml_weights_registry"
+path = "src/lib.rs"
+
+# External
+[dependencies]
+anyhow = "1.0"
+const_format = "0.2.34"

crates/desktop/ml_weights_registry/src/lib.rs

@@ -0,0 +1 @@
+pub mod weights_registry;

crates/desktop/ml_weights_registry/src/weights_registry/flux_weights.rs

@@ -0,0 +1,80 @@
+use crate::weights_registry::weight_descriptor::{WeightDescriptor, WeightFunction};
+use crate::weights_registry::weight_descriptor_builder::weight;
+
+
+/// https://huggingface.co/black-forest-labs/FLUX.1-dev
+pub const FLUX_DEV : WeightDescriptor = weight!(
+  "Flux-1 Dev",
+  "flux1-dev.safetensors",
+  WeightFunction::ImageGeneration
+);
+
+// TODO(bt): I think the autoencoder for Dev and Schnell are the same.
+//  They have the same checksum unless I mixed up the files when downloading.
+
+/// https://huggingface.co/black-forest-labs/FLUX.1-dev
+pub const FLUX_DEV_AUTOENCODER : WeightDescriptor = weight!(
+  "Flux-1 Dev (Autoencoder)",
+  "flux1-dev.ae.safetensors",
+  WeightFunction::ImageGeneration
+);
+
+/// https://huggingface.co/black-forest-labs/FLUX.1-schnell
+pub const FLUX_SCHNELL : WeightDescriptor = weight!(
+  "Flux-1 Schnell",
+  "flux1-schnell.safetensors",
+  WeightFunction::ImageGeneration
+);
+
+/// https://huggingface.co/black-forest-labs/FLUX.1-schnell
+pub const FLUX_SCHNELL_AUTOENCODER : WeightDescriptor = weight!(
+  "Flux-1 Schnell (Autoencoder)",
+  "flux1-schnell.ae.safetensors",
+  WeightFunction::ImageGeneration
+);
+
+/// https://huggingface.co/lmz/candle-flux
+pub const LMZ_CANDLE_FLUX_SCHNELL_QUANTIZED_GGUF : WeightDescriptor = weight!(
+  "LMZ Flux Schnell Quantized",
+  "lmz_candle_flux_flux1-schnell.gguf",
+  WeightFunction::ImageGeneration
+);
+
+/// https://huggingface.co/google/t5-v1_1-xxl/tree/main
+/// NB: I couldn't find the safetensors file here
+/// NB: This should live under "refs/pr/2"
+pub const GOOGLE_T5_V1_1_XXL_MODEL : WeightDescriptor = weight!(
+  "Google T5 v1.1 (XXL)",
+  "google_t5_v1_1_xxl_model.safetensors",
+  WeightFunction::TextEncoder
+);
+
+/// https://huggingface.co/google/t5-v1_1-xxl/tree/main
+/// NB: I couldn't find the safetensors file here
+/// NB: This should live under "refs/pr/2"
+pub const GOOGLE_T5_V1_1_XXL_CONFIG : WeightDescriptor = weight!(
+  "Google T5 v1.1 (XXL) Config",
+  "google_t5_v1_1_xxl_config.json",
+  WeightFunction::TextEncoder
+);
+
+/// https://huggingface.co/lmz/mt5-tokenizers
+pub const LMZ_T5_TOKENIZER_JSON : WeightDescriptor = weight!(
+  "LMZ T5 Tokenizer",
+  "lmz_mt5_tokenizers_t5-v1_1-xxl.tokenizer.json",
+  WeightFunction::TextTokenizer
+);
+
+/// https://huggingface.co/openai/clip-vit-large-patch14/tree/main
+pub const OPENAI_CLIP_VIT_P14_MODEL : WeightDescriptor = weight!(
+  "OpenAI CLIP VIT (patch 14) model",
+  "openai_clip-vit-large-patch14.model.safetensors",
+  WeightFunction::TextTokenizer
+);
+
+/// https://huggingface.co/openai/clip-vit-large-patch14/tree/main
+pub const OPENAI_CLIP_VIT_P14_TOKENIZER_JSON: WeightDescriptor = weight!(
+  "OpenAI CLIP VIT (patch 14) tokenizer",
+  "openai_clip-vit-large-patch14.tokenizer.json",
+  WeightFunction::TextTokenizer
+);

crates/desktop/ml_weights_registry/src/weights_registry/mod.rs

@@ -0,0 +1,4 @@
+mod weight_descriptor_builder;
+pub mod flux_weights;
+pub mod weight_descriptor;
+pub mod weights;

crates/desktop/ml_weights_registry/src/weights_registry/weight_descriptor.rs

@@ -0,0 +1,47 @@
+
+// TODO(bt,2025-03-13): When it's time for dynamic, non-const allocation, refer to
+//  http::header::name::HeaderName: https://github.com/hyperium/http/blob/master/src/header/name.rs
+//  as a pattern that will support both compile-time constants and dynamically allocated String-based variants.
+
+#[derive(Clone, Copy)]
+pub enum WeightFunction {
+  ImageGeneration,
+  ImageSegmentation,
+  TextEncoder,
+  TextTokenizer,
+  Unet,
+  Vae,
+}
+
+pub struct WeightDescriptor {
+  /// Name of the weight, eg. "Lykon Dreamshaper 7"
+  pub name: &'static str,
+
+  /// Name of the file on the filesystem, eg. lykon_dreamshaper_7_text_encoder.fp16.safetensors
+  pub filename: &'static str,
+
+  /// URL of the file in the R2 bucket,
+  /// eg. https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/lykon_dreamshaper_7_text_encoder.fp16.safetensors
+  pub r2_download_url: &'static str,
+
+  pub function: WeightFunction,
+}
+
+impl WeightFunction {
+  pub fn as_descriptive_str(&self) -> &'static str {
+    match self {
+      WeightFunction::ImageGeneration => "image generation",
+      WeightFunction::ImageSegmentation => "image segmentation",
+      WeightFunction::TextEncoder => "text encoder",
+      WeightFunction::TextTokenizer => "text tokenizer",
+      WeightFunction::Unet => "UNET",
+      WeightFunction::Vae => "VAE",
+    }
+  }
+}
+
+impl WeightDescriptor {
+  pub fn to_descriptive_name(&self) -> String {
+    format!("{} ({})", self.name, self.function.as_descriptive_str())
+  }
+}

crates/desktop/ml_weights_registry/src/weights_registry/weight_descriptor_builder.rs

@@ -0,0 +1,17 @@
+pub (super) const R2_BUCKET_URL : &str = "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/";
+
+// TODO(bt,2025-03-13): Try to use keyword args from the call site.
+// NB: We're using a macro because we can't use a builder pattern and then concat string literals
+//  and variables (even if they're known to be &'static) and keep allocations static.
+macro_rules! weight {
+    ($name: literal, $file_name:literal, $weight_function:expr) => {
+      $crate::weights_registry::weight_descriptor::WeightDescriptor {
+        name: $name,
+        filename: $file_name,
+        r2_download_url: const_format::concatcp!(crate::weights_registry::weight_descriptor_builder::R2_BUCKET_URL, $file_name),
+        function: $weight_function,
+      }
+    };
+}
+
+pub(crate) use weight;

crates/desktop/ml_weights_registry/src/weights_registry/weights.rs

@@ -0,0 +1,65 @@
+use crate::weights_registry::weight_descriptor::{WeightDescriptor, WeightFunction};
+use crate::weights_registry::weight_descriptor_builder::weight;
+
+pub const CLIP_JSON : WeightDescriptor = weight!(
+  "CLIP",
+  "clip_vit_bigg_14_lion2b_39b_b160k.tokenizer.json",
+  WeightFunction::TextTokenizer
+);
+
+/// For LCM-based image generation
+pub const LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16 : WeightDescriptor = weight!(
+  "Lykon Dreamshaper 7",
+  "lykon_dreamshaper_7_text_encoder.fp16.safetensors",
+  WeightFunction::TextEncoder
+);
+
+/// For LCM-based image generation
+pub const LYKON_DEAMSHAPER_7_VAE : WeightDescriptor = weight!(
+  "Lykon Dreamshaper 7",
+  "lykon_dreamshaper_7_vae.safetensors",
+  WeightFunction::Vae
+);
+
+/// For LCM-based image generation
+pub const SIMIANLUO_LCM_DREAMSHAPER_V7_UNET : WeightDescriptor = weight!(
+  "SimianLuo LCM Dreamshaper v7",
+  "simianluo_lcm_dreamshaper_v7_unet.safetensors",
+  WeightFunction::Unet
+);
+
+/// Image segmentation and background removal
+/// dis_medium.onnx
+pub const DIS_MEDIUM_ONNX : WeightDescriptor = weight!(
+  "Dichotomous Image Segmentation Medium",
+  "dis_medium.onnx",
+  WeightFunction::ImageSegmentation
+);
+
+#[deprecated]
+pub const SDXL_TURBO_CLIP_TEXT_ENCODER : WeightDescriptor = weight!(
+  "SDXL Turbo CLIP",
+  "clip_text_encoder.safetensors",
+  WeightFunction::TextEncoder
+);
+
+#[deprecated]
+pub const SDXL_TURBO_CLIP_TEXT_ENCODER_2 : WeightDescriptor = weight!(
+  "SDXL Turbo CLIP #2",
+  "clip_text_encoder_2.safetensors",
+  WeightFunction::TextEncoder
+);
+
+#[deprecated]
+pub const SDXL_TURBO_VAE : WeightDescriptor = weight!(
+  "SDXL Turbo",
+  "diffusion_pytorch_model.vae.safetensors",
+  WeightFunction::Vae
+);
+
+#[deprecated]
+pub const SDXL_TURBO_UNET : WeightDescriptor = weight!(
+  "SDXL Turbo",
+  "diffusion_pytorch_model.unet.safetensors",
+  WeightFunction::Unet
+);

crates/desktop/src-tauri/Cargo.toml

@@ -1,65 +0,0 @@
-[package]
-name = "app"
-version = "0.1.0"
-description = "A Tauri App"
-authors = ["you"]
-license = ""
-repository = ""
-edition = "2021"
-rust-version = "1.77.2"
-
-# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[lib]
-name = "app_lib"
-crate-type = ["staticlib", "cdylib", "rlib"]
-
-[build-dependencies]
-tauri-build = { version = "2.0.4", features = [] }
-
-[dependencies]
-
-# Internal
-filesys = { path = "../../lib/files/filesys" }
-memory_caching = { path = "../../lib/caching/memory_caching" }
-
-# Bypass cudarc issues on Windows https://github.com/coreylowman/cudarc/issues/332
-cudarc = "0.13.7"
-
-# External
-anyhow = "1.0"
-bytes = "1.10"
-base64 = "0.22.1"
-candle-core = { path = "../../vendor/candle/candle-core", features = ["cuda"] }
-candle-nn = { path = "../../vendor/candle/candle-nn", features = ["cuda"] }
-candle-transformers = { path = "../../vendor/candle/candle-transformers", features = ["cuda"] }
-hf-hub = "0.3.2"
-image = "0.25.5"
-serde_yaml = "0.9"
-log = "0.4"
-once_cell = "1.20.3"
-expanduser = "1.2.2"
-rand = "0.8"
-reqwest = "0.12.12"
-serde = { version = "1.0", features = ["derive"] }
-serde_json = "1.0"
-tauri = { version = "2.2.4", features = [] }
-tauri-plugin-log = "2"
-tokenizers = "0.15.2"
-
-[features]
-default = ["cuda"]
-cuda = ["candle-core/cuda", "candle-nn/cuda", "candle-transformers/cuda", "candle-transformers/flash-attn", "candle-core/cudnn"]
-
-[dev-dependencies]
-libloading = "0.8.6"
-
-[profile.dev]
-incremental = true # Compile your binary in smaller steps.
-
-[profile.release]
-codegen-units = 1 # Allows LLVM to perform better optimization.
-lto = true # Enables link-time-optimizations.
-opt-level = 3 # `s` prioritizes small binary size. Use `3` if you prefer speed.
-panic = "abort" # Higher performance by disabling panic handlers.
-strip = true # Ensures debug symbols are removed.

crates/desktop/src-tauri/src/endpoints/download_models.rs

@@ -1,16 +0,0 @@
-use crate::ml::downloads::download_all_models::download_all_models;
-use crate::state::app_dir::AppDataRoot;
-use tauri::State;
-
-#[tauri::command]
-pub async fn download_models(
-  app_data_root: State<'_, AppDataRoot>,
-) -> Result<String, String> {
-  
-  download_all_models(&app_data_root).await
-    .map_err(|err| {
-      "error downloading models".to_string()
-    })?;
-  
-  Ok(("downloaded".to_string()))
-}

crates/desktop/src-tauri/src/endpoints/image_endpoint.rs

@@ -1,143 +0,0 @@
-use crate::ml::downloads::download_all_models::download_all_models;
-use crate::ml::model_cache::ModelCache;
-use crate::ml::prompt_cache::PromptCache;
-use crate::ml::stable_diffusion::lcm_pipeline::{lcm_pipeline, Args};
-use crate::state::app_config::AppConfig;
-use base64::{engine::general_purpose::STANDARD as BASE64_STANDARD, Engine};
-use bytes::BytesMut;
-use image::imageops::FilterType;
-use image::{DynamicImage, ImageFormat, ImageReader};
-use log::{error, info};
-use std::io::Cursor;
-use std::path::PathBuf;
-use tauri::{AppHandle, State};
-use crate::state::app_dir::AppDataRoot;
-
-const PROMPT_FILENAME : &str = "prompt.txt";
-const PROMPT: &str = "A beautiful landscape with mountains and a lake";
-
-const NEGATIVE_PROMPT: &str = "bad quality, bad faces, poor quality, blurry faces, watermark";
-
-// TODO: Mark async https://github.com/tauri-apps/tauri/discussions/7737
-
-
-/// This handler takes an image (as a base64 encoded string) and a prompt and returns
-/// an image (as a base64-encoded string).
-#[tauri::command]
-pub async fn infer_image(
-  image: &str,
-  prompt: Option<String>,
-  strength: Option<u8>,
-  model_config: State<'_, AppConfig>,
-  model_cache: State<'_, ModelCache>,
-  prompt_cache: State<'_, PromptCache>,
-  app_data_root: State<'_, AppDataRoot>,
-  app: AppHandle,
-) -> Result<String, String> {
-
-  let prompt = get_prompt_or_fallback(prompt);
-
-  info!("Prompt: {}", prompt);
-  info!("Strength: {:?}", strength);
-
-  let image = hydrate_base64_image(image)
-    .map_err(|err| format!("Couldn't hydrate image from base64: {}", err))?;
-
-  // TODO(bt): Move this to another endpoint
-  download_all_models(&app_data_root).await
-    .map_err(|err| {
-      error!("couldn't download models: {:?}", err);
-      "Couldn't download models".to_string()
-    })?;
-
-  let result = do_infer_image(
-    &prompt, 
-    image, 
-    strength, 
-    &model_config, 
-    &model_cache, 
-    prompt_cache, 
-    app, 
-    &app_data_root
-  ).await;
-  
-  if let Err(err) = result.as_deref() {
-    error!("There was an error: {:?}", err);
-  }
-  
-  result
-}
-
-async fn do_infer_image(
-  prompt: &str,
-  image: DynamicImage,
-  strength: Option<u8>,
-  config: &AppConfig,
-  model_cache: &ModelCache,
-  prompt_cache: State<'_, PromptCache>,
-  app: AppHandle,
-  app_data_root: &AppDataRoot,
-) -> Result<String, String> {
-
-  let args = Args {
-    image: &image,
-    prompt: prompt.to_string(),
-    uncond_prompt: NEGATIVE_PROMPT.to_string(),
-    model_cache,
-    configs: config,
-    prompt_cache: &prompt_cache,
-    i2i_strength: strength,
-    cfg_scale: config.cfg_scale,
-    app: &app,
-    app_data_root,
-  };
-
-  match lcm_pipeline(args) {
-    Ok(image) => {
-      let mut bytes = Vec::with_capacity(1024*1024);
-      
-      image.write_to(&mut Cursor::new(&mut bytes), ImageFormat::Png)
-        .map_err(|err| format!("failure to encode image: {:?}", err))?;
-
-      let bytes = BASE64_STANDARD.encode(bytes);
-
-      Ok(bytes)
-    },
-    Err(e) => Err(format!("Failed to generate image: {}", e)),
-  }
-}
-
-fn get_prompt_or_fallback(user_prompt: Option<String>) -> String {
-  let user_prompt = user_prompt.map(|prompt| prompt.trim().to_string())
-    .filter(|prompt| !prompt.is_empty());
-
-  if let Some(prompt) = user_prompt {
-    return prompt;
-  }
-
-  let prompt_file = PathBuf::from(PROMPT_FILENAME)
-    .canonicalize()
-    .unwrap_or_else(|_| PathBuf::from(PROMPT_FILENAME));
-
-  std::fs::read_to_string(&prompt_file)
-    .map_err(|err| format!("Failed to read prompt file: {}", err))
-    .unwrap_or_else(|_| {
-      error!("Failed to read prompt file: {:?}", prompt_file);
-      PROMPT.to_string()
-    })
-    .trim()
-    .to_string()
-}
-
-fn hydrate_base64_image(base64_image: &str) -> anyhow::Result<DynamicImage> {
-  let bytes = BASE64_STANDARD.decode(base64_image)?;
-
-  let image = ImageReader::new(Cursor::new(bytes))
-    .with_guessed_format()?
-    .decode()?;
-
-  // TODO(bt,2025-02-17): Running out of vram with full image buffer size
-  let image = image.resize(512, 512, FilterType::CatmullRom);
-
-  Ok(image)
-}

crates/desktop/src-tauri/src/endpoints/mod.rs

@@ -1,4 +0,0 @@
-pub mod download_models;
-pub mod flip_image;
-pub mod image_endpoint;
-pub mod test_counter;

crates/desktop/src-tauri/src/endpoints/test_counter.rs

@@ -1,24 +0,0 @@
-use std::sync::{Arc, RwLock};
-use once_cell::sync::Lazy;
-
-static COUNTER : Lazy<Arc<RwLock<u64>>> = Lazy::new(|| Arc::new(RwLock::new(0)));
-
-#[tauri::command]
-pub fn test_counter() -> u64 {
-  let value : u64;
-  {
-    match COUNTER.write() {
-      Ok(mut counter) => {
-        *counter += 1;
-        value = *counter;
-      },
-      Err(_e) => {
-        value = 0;
-      }
-    }
-  }
-
-  println!("I was invoked from JavaScript! {:?}", value);
-
-  value
-}

crates/desktop/src-tauri/src/events/notification_event.rs

@@ -1,22 +0,0 @@
-use serde::Serialize;
-
-
-#[derive(Clone, Copy, Serialize)]
-#[serde(rename_all = "snake_case")]
-pub enum NotificationModelType {
-  Unet,
-  Vae,
-}
-
-#[derive(Clone, Serialize)]
-#[serde(rename_all = "snake_case")]
-pub enum NotificationEvent<'a> {
-  ModelDownloadStarted {
-    model_name: &'a str, 
-    model_type: NotificationModelType,
-  },
-  ModelDownloadComplete {
-    model_name: &'a str,
-    model_type: NotificationModelType,
-  }
-}

crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs

@@ -1,15 +0,0 @@
-use crate::ml::downloads::download_model_registry_file::download_model_registry_file;
-use crate::ml::model_type::ModelType;
-use crate::state::app_dir::AppDataRoot;
-
-pub async fn download_all_models(app_data_root: &AppDataRoot) -> anyhow::Result<()> {
-  let weights_dir = app_data_root.weights_dir();
-  
-  download_model_registry_file(ModelType::ClipJson, weights_dir).await?;
-  download_model_registry_file(ModelType::SdxlTurboUnet, weights_dir).await?;
-  download_model_registry_file(ModelType::SdxlTurboVae, weights_dir).await?;
-  download_model_registry_file(ModelType::SdxlTurboClipEncoder, weights_dir).await?;
-  download_model_registry_file(ModelType::SdxlTurboClipEncoder2, weights_dir).await?;
-  
-  Ok(())
-}

crates/desktop/src-tauri/src/ml/downloads/download_model_registry_file.rs

@@ -1,30 +0,0 @@
-use crate::ml::model_type::ModelType;
-use crate::state::app_dir::AppWeightsDir;
-use log::info;
-use std::fs::File;
-use std::path::{Path, PathBuf};
-
-pub async fn download_model_registry_file(
-  model_type: ModelType,
-  weights_dir: &AppWeightsDir,
-) -> anyhow::Result<PathBuf> {
-  
-  let path = weights_dir.model_path(&model_type);
-
-  if path.exists() {
-    return Ok(path);
-  }
-
-  info!("downloading model: {:?} to {:?}", model_type, path);
-
-  let body = reqwest::get(model_type.get_download_url())
-    .await?
-    .bytes()
-    .await?;
-  
-  let _file = File::create(&path)?;
-  
-  std::fs::write(&path, body.as_ref())?;
-  
-  Ok(path)
-}

crates/desktop/src-tauri/src/ml/downloads/mod.rs

@@ -1,2 +0,0 @@
-pub mod download_model_registry_file;
-pub mod download_all_models;

crates/desktop/src-tauri/src/ml/mod.rs

@@ -1,8 +0,0 @@
-pub mod image;
-pub mod model_cache;
-pub mod model_file;
-pub mod models;
-pub mod prompt_cache;
-pub mod stable_diffusion;
-pub mod model_type;
-pub mod downloads;

crates/desktop/src-tauri/src/ml/model_file.rs

@@ -1,83 +0,0 @@
-#[cfg(feature = "accelerate")]
-extern crate accelerate_src;
-#[cfg(feature = "mkl")]
-extern crate intel_mkl_src;
-
-use candle_transformers::models::stable_diffusion;
-
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-pub enum ModelFile {
-  Tokenizer,
-  Tokenizer2,
-  Clip,
-  Clip2,
-  Unet,
-  Vae,
-}
-
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-pub enum StableDiffusionVersion {
-  V1_5,
-  V1_5Inpaint,
-  V2_1,
-  V2Inpaint,
-  Xl,
-  XlInpaint,
-  Turbo,
-}
-
-impl StableDiffusionVersion {
-  pub fn repo(&self) -> &'static str {
-    match self {
-      Self::XlInpaint => "diffusers/stable-diffusion-xl-1.0-inpainting-0.1",
-      Self::Xl => "stabilityai/stable-diffusion-xl-base-1.0",
-      Self::V2Inpaint => "stabilityai/stable-diffusion-2-inpainting",
-      Self::V2_1 => "stabilityai/stable-diffusion-2-1",
-      Self::V1_5 => "runwayml/stable-diffusion-v1-5",
-      Self::V1_5Inpaint => "stable-diffusion-v1-5/stable-diffusion-inpainting",
-      Self::Turbo => "stabilityai/sdxl-turbo",
-    }
-  }
-}
-
-impl ModelFile {
-  pub fn get(
-    &self,
-    filename: Option<String>,
-    version: StableDiffusionVersion,
-    use_f16: bool,
-  ) -> anyhow::Result<std::path::PathBuf> {
-    use hf_hub::api::sync::Api;
-    match filename {
-      Some(filename) => Ok(std::path::PathBuf::from(filename)),
-      None => {
-        let (repo, path) = match self {
-          Self::Tokenizer => {
-            let tokenizer_repo = match version {
-              StableDiffusionVersion::V1_5
-              | StableDiffusionVersion::V2_1
-              | StableDiffusionVersion::V1_5Inpaint
-              | StableDiffusionVersion::V2Inpaint => "openai/clip-vit-base-patch32",
-              StableDiffusionVersion::Xl
-              | StableDiffusionVersion::XlInpaint
-              | StableDiffusionVersion::Turbo => {
-                "openai/clip-vit-large-patch14"
-              }
-            };
-            (tokenizer_repo, "tokenizer.json")
-          }
-          Self::Tokenizer2 => {
-            ("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", "tokenizer.json")
-          }
-          Self::Clip => (version.repo(), "text_encoder/model.safetensors"),
-          Self::Clip2 => (version.repo(), "text_encoder_2/model.safetensors"),
-          Self::Unet => (version.repo(), "unet/diffusion_pytorch_model.safetensors"),
-          Self::Vae => (version.repo(), "vae/diffusion_pytorch_model.safetensors"),
-        };
-        let filename = Api::new()?.model(repo.to_string()).get(path)?;
-        Ok(filename)
-      }
-    }
-  }
-}
-

crates/desktop/src-tauri/src/ml/model_type.rs

@@ -1,47 +0,0 @@
-use std::path::PathBuf;
-use crate::state::app_dir::AppWeightsDir;
-
-#[derive(Copy, Clone, Debug)]
-pub enum ModelType {
-  ClipJson,
-  SdxlTurboUnet,
-  SdxlTurboVae,
-  SdxlTurboClipEncoder,
-  SdxlTurboClipEncoder2,
-}
-
-impl ModelType {
-  pub fn get_hf_id(&self) -> Option<&'static str> {
-    match self {
-      Self::ClipJson => Some("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"),
-      Self::SdxlTurboUnet 
-      | Self::SdxlTurboVae
-      | Self::SdxlTurboClipEncoder
-      | Self::SdxlTurboClipEncoder2 => Some("stabilityai/sdxl-turbo"),
-    }
-  }
-
-  pub fn get_download_url(&self) -> &'static str {
-    match self {
-      Self::ClipJson => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/tokenizer.json",
-      Self::SdxlTurboUnet => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/diffusion_pytorch_model.unet.safetensors",
-      Self::SdxlTurboVae => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/diffusion_pytorch_model.vae.safetensors",
-      Self::SdxlTurboClipEncoder => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/clip_text_encoder.safetensors",
-      Self::SdxlTurboClipEncoder2 => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/clip_text_encoder_2.safetensors",
-    }
-  }
-
-  pub fn get_filename(&self) -> &'static str {
-    match self {
-      Self::ClipJson => "clip_vit_bigg_14_lion2b_39b_b160k.tokenizer.json",
-      Self::SdxlTurboUnet => "diffusion_pytorch_model.unet.safetensors",
-      Self::SdxlTurboVae => "diffusion_pytorch_model.vae.safetensors",
-      Self::SdxlTurboClipEncoder => "clip_text_encoder.safetensors",
-      Self::SdxlTurboClipEncoder2 => "clip_text_encoder_2.safetensors",
-    }
-  }
-  
-  pub fn get_path(&self, weights_dir: &AppWeightsDir) -> PathBuf {
-    weights_dir.model_path(self)
-  }
-}

crates/desktop/src-tauri/src/ml/models/lazy_load_vae_model.rs

@@ -1,130 +0,0 @@
-use crate::ml::model_file::StableDiffusionVersion;
-use anyhow::anyhow;
-use candle_core::{DType, Device, Tensor};
-use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
-use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
-use hf_hub::api::sync::Api;
-use std::sync::{Arc, RwLock};
-
-pub struct LazyLoadVaeModel {
-  lazy_model: Arc<RwLock<Option<AutoEncoderKL>>>,
-  model_configs: VaeConfigs,
-}
-
-// TODO: Maybe we pass these via a central config read at startup, passed as an Arc<ConfigT>
-pub struct VaeConfigs {
-  sd_config: StableDiffusionConfig,
-  sd_version: StableDiffusionVersion,
-  device: Device,
-  dtype: DType,
-  hf_api: Api,
-}
-
-impl LazyLoadVaeModel {
-  pub fn new(
-    sd_config: &StableDiffusionConfig,
-    sd_version: &StableDiffusionVersion,
-    hf_api: &Api,
-    device: &Device,
-    dtype: DType,
-  ) -> anyhow::Result<Self> {
-    println!("building vae model... (1)");
-    
-    Ok(Self {
-      lazy_model: Arc::new(RwLock::new(None)),
-      model_configs: VaeConfigs {
-        sd_config: sd_config.clone(),
-        sd_version: sd_version.clone(),
-        device: device.clone(),
-        hf_api: hf_api.clone(),
-        dtype
-      }
-    })
-  }
-
-  pub fn encode(&self, xs: &Tensor) -> anyhow::Result<DiagonalGaussianDistribution> {
-    println!("vae encode...");
-    match self.try_encode(xs) {
-      Err(err) => Err(err),
-      Ok(Some(result)) => Ok(result),
-      Ok(None) => {
-        // Model wasn't previously loaded
-        self.load_model()?;
-        self.try_encode(xs)?
-          .ok_or(anyhow!("model was not previously loaded"))
-      }
-    }
-  }
-  
-  pub fn decode(&self, xs: &Tensor) -> anyhow::Result<Tensor> {
-    println!("vae decode...");
-    match self.try_decode(xs) {
-      Err(err) => Err(err),
-      Ok(Some(result)) => Ok(result),
-      Ok(None) => {
-        // Model wasn't previously loaded
-        self.load_model()?;
-        self.try_decode(xs)?
-          .ok_or(anyhow!("model was not previously loaded"))
-      }
-    }
-  }
-
-  fn try_encode(&self, xs: &Tensor) -> anyhow::Result<Option<DiagonalGaussianDistribution>> {
-    match self.lazy_model.read() {
-      Err(err) => Err(anyhow!("lock error: {:?}", err)),
-      Ok(model) => {
-        match &*model {
-          None => Ok(None), // Model is not yet loaded
-          Some(model) => {
-            Ok(Some(model.encode(xs)?))
-          }
-        }
-      }
-    }
-  }
-
-  fn try_decode(&self, xs: &Tensor) -> anyhow::Result<Option<Tensor>> {
-    match self.lazy_model.read() {
-      Err(err) => Err(anyhow!("lock error: {:?}", err)),
-      Ok(model) => {
-        match &*model {
-          None => Ok(None), // Model is not yet loaded
-          Some(model) => {
-            Ok(Some(model.decode(xs)?))
-          }
-        }
-      }
-    }
-  }
-
-  fn load_model(&self) -> anyhow::Result<()> {
-    println!("load vae");
-    match self.lazy_model.write() {
-      Err(err) => return Err(anyhow!("lock error: {:?}", err)),
-      Ok(mut maybe_model) => {
-        match &*maybe_model {
-          Some(_model) => {} // Already loaded; fall through
-          None => {
-            let repo = self.model_configs.sd_version.repo();
-
-            println!("Building VAE model from : {:?} ... (3)", repo);
-            
-            let vae_file = self.model_configs.hf_api.model(repo.to_string())
-              .get("vae/diffusion_pytorch_model.safetensors")?;
-
-            println!("Building VAE model from file {:?}...", &vae_file);
-            let vae = self.model_configs
-              .sd_config
-              .build_vae(vae_file, &self.model_configs.device, self.model_configs.dtype)?;
-
-            println!("Built VAE...");
-            *maybe_model = Some(vae);
-          }
-        }
-      }
-    }
-    Ok(())
-  }
- 
-}

crates/desktop/src-tauri/src/ml/models/mod.rs

@@ -1,2 +0,0 @@
-pub mod unet_model;
-pub mod lazy_load_vae_model;

crates/desktop/src-tauri/src/ml/models/unet_model.rs

@@ -1,26 +0,0 @@
-use candle_core::{DType, Device, Tensor};
-use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModel;
-use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
-use std::path::Path;
-
-pub struct UNetModel {
-  model: UNet2DConditionModel,
-}
-
-impl UNetModel {
-  pub fn new<P: AsRef<Path>>(sd_config: &StableDiffusionConfig, unet_weights_file: P, device: &Device, dtype: DType) -> anyhow::Result<Self> {
-    println!("building unet model... (1)");
-
-    let model = sd_config.build_unet(unet_weights_file, &device, 4, true, dtype)?;
-
-    Ok(Self { model })
-  }
-
-  pub fn inference(&self, latent_model_input: &Tensor, timestep: f64, text_embeddings: &Tensor) -> anyhow::Result<Tensor> {
-    Ok(self.model.forward(&latent_model_input, timestep, &text_embeddings)?)
-  }
-
-  pub fn forward_with_guidance(&self, latent_model_input: &Tensor, timestep: f64, text_embeddings: &Tensor, guidance_scale_embedding: &Tensor) -> anyhow::Result<Tensor> {
-    Ok(self.model.forward_with_guidance(&latent_model_input, timestep, &text_embeddings, Some(guidance_scale_embedding))?)
-  }
-}

crates/desktop/src-tauri/src/ml/stable_diffusion/infer_clip_text_embeddings.rs

@@ -1,170 +0,0 @@
-#[cfg(feature = "accelerate")]
-extern crate accelerate_src;
-#[cfg(feature = "mkl")]
-extern crate intel_mkl_src;
-
-use std::path::PathBuf;
-use candle_transformers::models::stable_diffusion;
-
-use crate::ml::model_file::{ModelFile, StableDiffusionVersion};
-use anyhow::Error as E;
-use candle_core::{DType, Device, Module, Tensor, D};
-use log::info;
-use tokenizers::Tokenizer;
-use crate::ml::model_type::ModelType;
-use crate::state::app_dir::AppWeightsDir;
-
-#[allow(clippy::too_many_arguments)]
-pub fn infer_clip_text_embeddings(
-  prompt: &str,
-  uncond_prompt: &str,
-  tokenizer: Option<String>,
-  clip_weights: Option<String>,
-  clip2_weights: Option<String>,
-  sd_version: StableDiffusionVersion,
-  sd_config: &stable_diffusion::StableDiffusionConfig,
-  use_f16: bool,
-  device: &Device,
-  dtype: DType,
-  use_guide_scale: bool,
-  weights_dir: &AppWeightsDir,
-) -> anyhow::Result<Tensor> {
-  info!("Preparing text embeddings...");
-
-  let which = match sd_version {
-    StableDiffusionVersion::Xl
-    | StableDiffusionVersion::XlInpaint
-    | StableDiffusionVersion::Turbo => vec![true, false],
-    _ => vec![true],
-  };
-
-  let text_embeddings = which
-    .iter()
-    .map(|first| {
-      do_infer_clip_text_embeddings(
-        &prompt,
-        &uncond_prompt,
-        None, // tokenizer
-        None, // clip_weights
-        None, // clip2_weights
-        sd_version,
-        &sd_config,
-        false, // use_f16
-        &device,
-        dtype,
-        use_guide_scale,
-        *first,
-        weights_dir,
-      )
-    })
-    .collect::<anyhow::Result<Vec<_>>>()?;
-
-  let text_embeddings = Tensor::cat(&text_embeddings, D::Minus1)?;
-  
-  Ok(text_embeddings)
-}
-
-#[allow(clippy::too_many_arguments)]
-fn do_infer_clip_text_embeddings(
-    prompt: &str,
-    uncond_prompt: &str,
-    tokenizer: Option<String>,
-    clip_weights: Option<String>,
-    clip2_weights: Option<String>,
-    sd_version: StableDiffusionVersion,
-    sd_config: &stable_diffusion::StableDiffusionConfig,
-    use_f16: bool,
-    device: &Device,
-    dtype: DType,
-    use_guide_scale: bool,
-    first: bool,
-    weights_dir: &AppWeightsDir,
-  ) -> anyhow::Result<Tensor> {
-  
-  let tokenizer_file = if first {
-    info!("ModelFile::Tokenizer");
-    ModelFile::Tokenizer
-  } else {
-    info!("ModelFile::Tokenizer2");
-    ModelFile::Tokenizer2
-  };
-  
-  info!("Loading clip from download");
-  //let tokenizer = tokenizer_file.get(tokenizer, sd_version, use_f16)?;
-  let tokenizer = weights_dir.model_path(&ModelType::ClipJson);
-  
-  info!("Tokenizer path: {:?}", tokenizer);
-  
-  let tokenizer = Tokenizer::from_file(tokenizer).map_err(E::msg)?;
-  
-  let pad_id = match &sd_config.clip.pad_with {
-    Some(padding) => *tokenizer.get_vocab(true).get(padding.as_str()).unwrap(),
-    None => *tokenizer.get_vocab(true).get("<|endoftext|>").unwrap(),
-  };
-
-  println!("Running with prompt \"{prompt}\".");
-  let mut tokens = tokenizer
-    .encode(prompt, true)
-    .map_err(E::msg)?
-    .get_ids()
-    .to_vec();
-  if tokens.len() > sd_config.clip.max_position_embeddings {
-    anyhow::bail!(
-            "the prompt is too long, {} > max-tokens ({})",
-            tokens.len(),
-            sd_config.clip.max_position_embeddings
-        )
-  }
-  while tokens.len() < sd_config.clip.max_position_embeddings {
-    tokens.push(pad_id)
-  }
-  let tokens = Tensor::new(tokens.as_slice(), device)?.unsqueeze(0)?;
-
-  println!("Building the Clip transformer.");
-  let clip_weights_file = if first {
-    ModelFile::Clip
-  } else {
-    ModelFile::Clip2
-  };
-  let clip_weights = if first {
-    //clip_weights_file.get(clip_weights, sd_version, use_f16)?
-    weights_dir.model_path(&ModelType::SdxlTurboClipEncoder)
-  } else {
-    //clip_weights_file.get(clip2_weights, sd_version, use_f16)?
-    weights_dir.model_path(&ModelType::SdxlTurboClipEncoder2)
-  };
-  let clip_config = if first {
-    &sd_config.clip
-  } else {
-    sd_config.clip2.as_ref().unwrap()
-  };
-  let text_model = stable_diffusion::build_clip_transformer(clip_config, clip_weights, device, DType::F32)?;
-
-  let text_embeddings = text_model.forward(&tokens)?;
-
-  let text_embeddings = if use_guide_scale {
-    let mut uncond_tokens = tokenizer
-      .encode(uncond_prompt, true)
-      .map_err(E::msg)?
-      .get_ids()
-      .to_vec();
-    if uncond_tokens.len() > sd_config.clip.max_position_embeddings {
-      anyhow::bail!(
-                "the negative prompt is too long, {} > max-tokens ({})",
-                uncond_tokens.len(),
-                sd_config.clip.max_position_embeddings
-            )
-    }
-    while uncond_tokens.len() < sd_config.clip.max_position_embeddings {
-      uncond_tokens.push(pad_id)
-    }
-
-    let uncond_tokens = Tensor::new(uncond_tokens.as_slice(), device)?.unsqueeze(0)?;
-    let uncond_embeddings = text_model.forward(&uncond_tokens)?;
-
-    Tensor::cat(&[uncond_embeddings, text_embeddings], 0)?.to_dtype(dtype)?
-  } else {
-    text_embeddings.to_dtype(dtype)?
-  };
-  Ok(text_embeddings)
-}

crates/desktop/src-tauri/src/ml/stable_diffusion/lcm_pipeline.rs

@@ -1,293 +0,0 @@
-use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
-
-use crate::events::notification_event::{NotificationModelType, NotificationEvent};
-use crate::ml::image::dynamic_image_to_tensor::dynamic_image_to_tensor;
-use crate::ml::image::tensor_to_image_buffer::{tensor_to_image_buffer, RgbImage};
-use crate::ml::model_cache::ModelCache;
-use crate::ml::model_file::StableDiffusionVersion;
-use crate::ml::model_type::ModelType;
-use crate::ml::models::unet_model::UNetModel;
-use crate::ml::prompt_cache::PromptCache;
-use crate::ml::stable_diffusion::get_vae_scale::get_vae_scale;
-use crate::ml::stable_diffusion::infer_clip_text_embeddings::infer_clip_text_embeddings;
-use crate::state::app_config::AppConfig;
-use anyhow::{anyhow, Error as E, Result};
-use candle_core::{DType, IndexOp, Tensor, D};
-use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
-use candle_transformers::models::stable_diffusion::lcm::LCMScheduler;
-use image::DynamicImage;
-use log::info;
-use rand::Rng;
-use tauri::{AppHandle, Emitter};
-use crate::state::app_dir::AppDataRoot;
-
-pub struct Args<'a> {
-    pub image: &'a DynamicImage,
-    pub prompt: String,
-    pub uncond_prompt: String,
-    pub cfg_scale: Option<f64>,
-    pub i2i_strength: Option<u8>,
-    pub configs: &'a AppConfig,
-    pub model_cache: &'a ModelCache,
-    pub prompt_cache: &'a PromptCache,
-    pub app_data_root: &'a AppDataRoot,
-    pub app: &'a AppHandle,
-}
-
-pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
-    let Args { 
-        prompt, 
-        uncond_prompt, 
-        cfg_scale, 
-        i2i_strength, 
-        configs, 
-        model_cache, 
-        prompt_cache, 
-        app, 
-        image,  
-        app_data_root,
-    } = args;
-
-    let weights_dir = app_data_root.weights_dir();
-    
-    println!("Starting image generation with the following configuration:");
-    println!("  Model: {:?}", configs.sd_version);
-    println!("  Prompt: {}", prompt);
-    println!("  Steps: {}", configs.scheduler_steps);
-    println!("  Device: {:?}", configs.device);
-
-    println!("Model dimensions: {}x{}", configs.sd_config.width, configs.sd_config.height);
-
-    // Use LCM Scheduler instead of Euler Ancestral for better speed and quality
-    let mut scheduler = LCMScheduler::new(
-        configs.scheduler_steps,
-        candle_transformers::models::stable_diffusion::lcm::LCMSchedulerConfig::default(),
-    )?;
-
-    let seed = configs.seed.unwrap_or_else(|| rand::thread_rng().gen());
-    configs.device.set_seed(seed)?;
-
-    info!("Using seed: {}", seed);
-
-    let guidance_scale = match cfg_scale {
-        Some(guidance_scale) => guidance_scale,
-        None => match configs.sd_version {
-            StableDiffusionVersion::V1_5
-            | StableDiffusionVersion::V2_1
-            | StableDiffusionVersion::Xl => 7.5,
-            StableDiffusionVersion::Turbo => 0.,
-            _ => 0., // NB: Not sure what the other model families should use, so sticking with "0"
-        },
-    };
-
-    let use_guide_scale = guidance_scale > 1.0;
-
-    info!("Using guide scale: {}", use_guide_scale);
-
-    info!("Checking if prompt is cached");
-    let maybe_cached = prompt_cache.get_copy(&prompt)?;
-
-    let mut text_embeddings = if let Some(tensor) = maybe_cached {
-        tensor
-    } else {
-        
-        info!("Prompt is NOT cached! Calculating embedding...");
-        let tensor = infer_clip_text_embeddings(
-            &prompt,
-            &uncond_prompt,
-            None, // tokenizer
-            None, // clip_weights
-            None, // clip2_weights
-            configs.sd_version,
-            &configs.sd_config,
-            false, // use_f16
-            &configs.device,
-            configs.dtype,
-            use_guide_scale,
-            weights_dir,
-        )?;
-        prompt_cache.store_copy(&prompt, &tensor)?;
-        tensor
-    };
-
-    println!("Text embeddings shape: {:?}", text_embeddings.shape());
-
-    println!("Loading input image into tensor...");
-
-    let input_image = dynamic_image_to_tensor(
-        image,
-        &configs.device,
-        configs.dtype)?;
-
-    println!("Reference image shape: {:?}", input_image.shape());
-
-    let vae_scale = get_vae_scale(configs.sd_version);
-
-    let maybe_vae = model_cache.get_vae()?;
-    
-    let vae = match maybe_vae {
-        Some(vae) => vae,
-        None => {
-            info!("No vae found in cache; loading...");
-            
-            let mut notify_download_complete = false;
-            //if !vae_file.exists() {
-            if true {
-                notify_download_complete = true;
-                app.emit("notification", NotificationEvent::ModelDownloadStarted {
-                    model_name: "sdxl-turbo-vae",
-                    model_type: NotificationModelType::Vae,
-                })?;
-            }
-
-            let vae_file = weights_dir.model_path(&ModelType::SdxlTurboVae);
-            
-            let vae = configs
-              .sd_config
-              .build_vae(vae_file, &configs.device, configs.dtype)?;
-
-            let vae = Arc::new(vae);
-            
-            model_cache.set_vae(vae.clone())?;
-            
-            if notify_download_complete {
-                app.emit("notification", NotificationEvent::ModelDownloadComplete {
-                    model_name: "sdxl-turbo-vae",
-                    model_type: NotificationModelType::Vae,
-                })?;
-            }
-            
-            vae
-        }
-    };
-
-    let maybe_unet = model_cache.get_unet()?;
-
-    let unet = match maybe_unet {
-        Some(unet) => unet,
-        None => {
-            info!("No unet found in cache; loading...");
-
-            let mut notify_download_complete = false;
-            //if !unet_file.exists() {
-            if true {
-                notify_download_complete = true;
-                app.emit("notification", NotificationEvent::ModelDownloadStarted {
-                    model_name: "sdxl-turbo-unet",
-                    model_type: NotificationModelType::Unet,
-                })?;
-            }
-            
-            let unet_file = weights_dir.model_path(&ModelType::SdxlTurboUnet);
-
-            let unet = UNetModel::new(&configs.sd_config, unet_file, &configs.device, configs.dtype)
-              .map_err(|err| anyhow!("error initializing unet model: {:?}", err))?;
-            
-            let unet = Arc::new(unet);
-
-            model_cache.set_unet(unet.clone())?;
-
-            if notify_download_complete {
-                app.emit("notification", NotificationEvent::ModelDownloadComplete {
-                    model_name: "sdxl-turbo-unet",
-                    model_type: NotificationModelType::Unet,
-                })?;
-            }
-
-            unet
-        }
-    };
-
-    let init_latent_dist : DiagonalGaussianDistribution = vae.encode(&input_image)?;
-
-    // Parse and calculate the img2img strength
-    let img2img_strength = match i2i_strength {
-        None => 0.75f64,
-        Some(strength) => (strength as f64) / 100.0f64,
-    };
-
-    // Create LCM guidance scale embeddings
-    let embedding_dim = 256;
-    let guidance_scale_embedding = scheduler.get_guidance_scale_embedding(
-        guidance_scale, 
-        embedding_dim, 
-        &configs.device, 
-        configs.dtype
-    )?;
-
-    // Generate latents from input image using the LCM approach
-    println!("Generating latents from input image...");
-    let init_latents = (init_latent_dist.sample()? * vae_scale)?;
-    println!("Initial latents shape: {:?}", init_latents.shape());
-
-    // Calculate timesteps for LCM with proper img2img handling
-    let timesteps = LCMScheduler::get_timesteps_for_steps(configs.scheduler_steps, img2img_strength);
-    let t_start = timesteps[0]; // First denoising step
-
-    // Add noise at the appropriate timestep for img2img
-    println!("Adding noise to latents for t_start={}", t_start);
-    let noise = init_latents.randn_like(0f64, 1f64)?;
-    let latents = scheduler.add_noise(&init_latents, noise, t_start)?;
-
-    println!("Latents initialized successfully");
-    println!("Initial noise shape: {:?}", latents.shape());
-
-    println!("Starting diffusion process...");
-    
-    let mut latents = latents;
-    let timesteps: Vec<_> = scheduler.timesteps().iter().copied().collect();
-
-    for (timestep_index, &timestep) in timesteps.iter().enumerate() {
-        println!("Processing step {}/{} @ timestamp = {}", timestep_index + 1, timesteps.len(), timestep);
-
-        let latent_model_input = if use_guide_scale {
-            Tensor::cat(&[&latents, &latents], 0)?
-        } else {
-            latents.clone()
-        };
-
-        let latent_model_input = scheduler.scale_model_input(latent_model_input, timestep);
-
-        // Use the guidance scale embedding in the UNet inference with the proper method name
-        let mut noise_pred = match unet.forward_with_guidance(
-            &latent_model_input, 
-            timestep as f64, 
-            &text_embeddings,
-            &guidance_scale_embedding) {
-            Ok(pred) => pred,
-            Err(e) => {
-                println!("UNet inference failed with error: {}", e);
-                return Err(anyhow::anyhow!("UNet inference failed: {}", e));
-            },
-        };
-
-        if use_guide_scale {
-            let chunks = noise_pred.chunk(2, 0)?;
-            let (noise_pred_uncond, noise_pred_text) = (&chunks[0], &chunks[1]);
-            noise_pred = (noise_pred_uncond + (guidance_scale * (noise_pred_text - noise_pred_uncond)?)?)?;
-        }
-
-        // Apply the LCM scheduler step
-        latents = scheduler.step(&noise_pred, timestep, &latents, timestep_index, configs.scheduler_steps)?;
-    }
-
-    println!("Diffusion process completed, decoding image...");
-    let image = vae.decode(&(latents / vae_scale)?)?;
-    
-    println!("VAE decode completed");
-
-    println!("Scaling image back");
-    let image = ((image / 2.)? + 0.5)?;
-    
-    println!("Normalized image values");
-    let image = (image.clamp(0f32, 1.)? * 255.)?;
-
-    println!("Convert to int8");
-    let image = image.to_dtype(DType::U8)?;
-    
-    println!("Converted to 8-bit format");
-
-    let image = tensor_to_image_buffer(&image.i(0)?)?;
-
-    Ok(image)
-}

crates/desktop/src-tauri/src/state/app_config.rs

@@ -1,108 +0,0 @@
-use crate::ml::model_file::StableDiffusionVersion;
-use crate::state::app_dir::AppDataRoot;
-use crate::state::os_platform::OsPlatform;
-use crate::state::yaml_config::YamlConfig;
-use candle_core::{DType, Device};
-use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
-use expanduser::expanduser;
-use filesys::recursively_find_file_by_name::recursively_find_file_by_name;
-use hf_hub::api::sync::Api;
-use log::info;
-use std::path::PathBuf;
-
-const DEFAULT_SD_IMAGE_WIDTH: usize = 512;
-const DEFAULT_SD_IMAGE_HEIGHT: usize = 512;
-
-const DEFAULT_SD_NUMERIC_DATATYPE: DType = DType::F32;
-const SD_VERSION: StableDiffusionVersion = StableDiffusionVersion::Turbo;
-
-const DEFAULT_DATA_DIR_UNIX : &str = "~/artcraft";
-
-/// A central place to configure app-wide details.
-pub struct AppConfig {
-  /// We only support one device type currently, and we can't multiplex 
-  /// over several devices yet.
-  pub device: Device,
-  
-  /// The core numeric type to use for models.
-  pub dtype: DType,
-  
-  pub sd_version: StableDiffusionVersion,
-  pub sd_config: StableDiffusionConfig,
-
-  pub image_height: usize,
-  pub image_width: usize,
-  pub scheduler_steps: usize,
-  pub scheduler_samples: usize,
-  
-  pub seed: Option<u64>,
-
-  pub cfg_scale: Option<f64>,
-  
-  /// Probably shouldn't be here.
-  pub hf_api: Api,
-  
-  /// Root location for application data.
-  pub app_data_root: AppDataRoot,
-}
-
-impl AppConfig {
-  
-  pub fn init() -> anyhow::Result<Self> {
-    let yaml_configs = YamlConfig::load_from_config_file_recursive();
-    
-    println!("Configs: {:?}", yaml_configs);
-    
-    let device = Device::new_cuda(0)
-      .unwrap_or_else(|e| {
-          info!("CUDA not available ({}), falling back to CPU", e);
-          Device::Cpu
-        });
-    
-    let sd_config = match SD_VERSION {
-      StableDiffusionVersion::Turbo => {
-        StableDiffusionConfig::sdxl_turbo(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
-      }
-      _ => {
-        StableDiffusionConfig::v2_1(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
-      }
-    };
-    
-    let hf_api = Api::new()?;
-    
-    let os_platform = OsPlatform::get();
-
-    let default_data_path = match os_platform {
-      OsPlatform::Linux => yaml_configs.linux_default_data_path.as_deref(),
-      OsPlatform::MacOs => yaml_configs.mac_default_data_path.as_deref(),
-      OsPlatform::Windows => yaml_configs.windows_default_data_path.as_deref(),
-    };
-    
-    let data_root_path = match default_data_path {
-      Some(default_data_path) => default_data_path.to_string(),
-      None => match os_platform { 
-        OsPlatform::Linux | OsPlatform::MacOs => DEFAULT_DATA_DIR_UNIX.to_string(),
-        OsPlatform::Windows => ".".to_string(),
-      }
-    };
-      
-    println!("Using data root: {}", data_root_path);
-
-    let app_data_root = AppDataRoot::create_existing(data_root_path)?;
-    
-    Ok(Self {
-      device,
-      dtype: DEFAULT_SD_NUMERIC_DATATYPE,
-      sd_version: SD_VERSION,
-      sd_config,
-      image_height: yaml_configs.image_height.unwrap_or(DEFAULT_SD_IMAGE_HEIGHT),
-      image_width: yaml_configs.image_width.unwrap_or(DEFAULT_SD_IMAGE_WIDTH),
-      scheduler_steps: yaml_configs.scheduler_steps.unwrap_or(4),
-      scheduler_samples: yaml_configs.scheduler_samples.unwrap_or(15),
-      seed: yaml_configs.seed,
-      cfg_scale: yaml_configs.cfg_scale,
-      hf_api,
-      app_data_root,
-    })
-  }
-}

crates/desktop/src-tauri/src/state/app_dir.rs

@@ -1,123 +0,0 @@
-use crate::state::os_platform::OsPlatform;
-use expanduser::expanduser;
-use std::path::{Path, PathBuf};
-use crate::ml::model_type::ModelType;
-
-const ASSETS_SUBDIRECTORY : &str = "assets";
-const WEIGHTS_SUBDIRECTORY : &str = "weights";
-
-/// The path to the application data directory, which includes "asset" and "weights" data.
-#[derive(Clone)]
-pub struct AppDataRoot {
-  path: PathBuf,
-  assets_dir: AppAssetsDir,
-  weights_dir: AppWeightsDir,
-}
-
-#[derive(Clone)]
-pub struct AppAssetsDir {
-  path: PathBuf,
-}
-
-#[derive(Clone)]
-pub struct AppWeightsDir {
-  path: PathBuf,
-}
-
-impl AppDataRoot {
-  pub fn create_existing<P: AsRef<Path>>(dir: P) -> anyhow::Result<Self> {
-    let mut dir = dir.as_ref().to_path_buf();
-    
-    match OsPlatform::get() {
-      OsPlatform::Linux | OsPlatform::MacOs => {
-        if let Some(d) = dir.as_os_str().to_str() {
-          dir = expanduser(d)?;
-        }
-      },
-      OsPlatform::Windows => {}
-    }
-    
-    if !dir.is_dir() {
-      println!("Creating directory {:?}", dir);
-      std::fs::create_dir_all(&dir)?;
-    }
-
-    match dir.canonicalize() {
-      Ok(d) => dir = d,
-      Err(err) => {
-        println!("Error canonicalizing {:?}: {}", dir, err);
-      }
-    }
-    
-    let assets_dir = AppAssetsDir::create_existing(dir.join(ASSETS_SUBDIRECTORY))?;
-    let weights_dir = AppWeightsDir::create_existing(dir.join(WEIGHTS_SUBDIRECTORY))?;
-    
-    Ok(Self {
-      path: dir,
-      assets_dir,
-      weights_dir,
-    })
-  }
-  
-  pub fn assets_dir(&self) -> &AppAssetsDir {
-    &self.assets_dir
-  }
-  
-  pub fn weights_dir(&self) -> &AppWeightsDir {
-    &self.weights_dir
-  }
-  
-  pub fn path(&self) -> &Path {
-    &self.path
-  }
-}
-
-impl AppAssetsDir {
-  pub fn create_existing<P: AsRef<Path>>(dir: P) -> anyhow::Result<Self> {
-    let mut dir = dir.as_ref().to_path_buf();
-    match dir.canonicalize() {
-      Ok(d) => dir = d,
-      Err(err) => {
-        println!("Error canonicalizing {:?}: {}", dir, err);
-      }
-    }
-    if !dir.exists() {
-      println!("Creating directory {:?}", dir);
-      std::fs::create_dir(&dir)?;
-    }
-    Ok(Self {
-      path: dir,
-    })
-  }
-
-  pub fn path(&self) -> &Path {
-    &self.path
-  }
-}
-
-impl AppWeightsDir {
-  pub fn create_existing<P: AsRef<Path>>(dir: P) -> anyhow::Result<Self> {
-    let mut dir = dir.as_ref().to_path_buf();
-    match dir.canonicalize() {
-      Ok(d) => dir = d,
-      Err(err) => {
-        println!("Error canonicalizing {:?}: {}", dir, err);
-      }
-    }
-    if !dir.exists() {
-      println!("Creating directory {:?}", dir);
-      std::fs::create_dir(&dir)?;
-    }
-    Ok(Self {
-      path: dir,
-    })
-  }
-
-  pub fn path(&self) -> &Path {
-    &self.path
-  }
-  
-  pub fn model_path(&self, model_registry: &ModelType) -> PathBuf {
-    self.path.join(model_registry.get_filename())
-  }
-}

crates/desktop/src-tauri/src/threads/downloader_thread.rs

@@ -1,11 +0,0 @@
-use tauri::{AppHandle, Emitter};
-use crate::state::app_dir::AppDataRoot;
-
-pub async fn downloader_thread(app_data_root: AppDataRoot, app: AppHandle) -> ! {
-  loop {
-    app.emit("testing", 1);
-    
-    println!(">>> Testing: {:?}", app_data_root.path());
-    std::thread::sleep(std::time::Duration::from_millis(500));
-  }
-}

crates/desktop/tauri-4o/.taurignore

@@ -0,0 +1,4 @@
+# Don't restart the dev server on Rust changes
+crates/
+Cargo.toml
+Cargo.lock

crates/desktop/tauri-4o/Cargo.toml

@@ -0,0 +1,74 @@
+[package]
+name = "artcraft"
+version = "0.1.0"
+description = "Artcraft"
+authors = ["you"]
+license = ""
+repository = ""
+edition = "2021"
+rust-version = "1.77.2"
+publish = false
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[lib]
+name = "app_lib"
+crate-type = ["staticlib", "cdylib", "rlib"]
+
+[build-dependencies]
+tauri-build = { version = "2.1.1", features = [] }
+
+[dependencies]
+# Workspace hack (faster compile times)
+my-workspace-hack = { workspace =  true }
+
+# Internal
+crockford = { workspace = true }
+errors = { workspace = true }
+filesys = { workspace = true }
+memory_caching = { workspace = true }
+openai_sora_client = { workspace = true }
+primitives = { workspace = true }
+subprocess_common = { workspace = true }
+
+# External
+#openssl = { version = "0.10", features = ["vendored"] }
+# NB(bt,2025-03-10): default-features = false on Windows while I debug openssl / rustls
+anyhow = { workspace = true }
+base64 = "0.22.1"
+bytes = "1.10"
+chrono = { workspace = true }
+const_format = "0.2.34"
+directories = "6.0.0"
+dirs = "6.0.0"
+futures = "0.3"
+image = "0.25.5"
+log = { workspace = true }
+once_cell = { workspace = true }
+rand = { workspace = true }
+reqwest = { workspace = true, features = ["multipart", "stream"], default-features = false }
+serde = { workspace = true }
+serde_derive = { workspace = true }
+serde_json = { workspace = true }
+serde_yaml = { workspace = true }
+tauri = { version = "2.4.1", features = ["unstable"] }
+tauri-plugin-log = "2.3.1"
+tempdir = "0.3.7"
+tempfile = "3.18.0"
+tokio = { version = "1.42", features = ["rt", "rt-multi-thread", "fs"] }
+tokio-util = { version = "0.7", features = ["codec"] }
+
+[dev-dependencies]
+libloading = "0.8.6"
+serial_test = "3.2.0"
+assertor = "0.0.3"
+
+[profile.dev]
+incremental = true # Compile your binary in smaller steps.
+
+[profile.release]
+codegen-units = 1 # Allows LLVM to perform better optimization.
+lto = true # Enables link-time-optimizations.
+opt-level = 3 # `s` prioritizes small binary size. Use `3` if you prefer speed.
+panic = "abort" # Higher performance by disabling panic handlers.
+strip = true # Ensures debug symbols are removed.

crates/desktop/tauri-4o/src/commands/image_generation_command.rs

@@ -0,0 +1,90 @@
+use std::fs::read_to_string;
+use base64::prelude::BASE64_STANDARD;
+use errors::AnyhowResult;
+use base64::Engine;
+use image::codecs::png::{CompressionType, FilterType, PngEncoder};
+use image::{DynamicImage, ImageReader};
+use std::io::Cursor;
+use log::{error, info};
+use tauri::{AppHandle, Manager, State};
+use openai_sora_client::credentials::SoraCredentials;
+use openai_sora_client::image_gen::common::{ImageSize, NumImages};
+use openai_sora_client::image_gen::sora_image_gen_remix::{sora_image_gen_remix, SoraImageGenRemixRequest};
+use openai_sora_client::upload::upload_media_from_bytes::sora_media_upload_from_bytes;
+use openai_sora_client::upload::upload_media_from_file::SoraMediaUploadRequest;
+use crate::state::app_dir::AppDataRoot;
+
+#[tauri::command]
+pub async fn image_generation_command(
+  image: &str,
+  prompt: &str,
+  app: AppHandle,
+  app_data_root: State<'_, AppDataRoot>,
+) -> Result<String, String> {
+  info!("image_generation_command called; processing image...");
+
+  let bytes = BASE64_STANDARD.decode(image)
+    .map_err(|err| format!("Base64 decode error: {}", err))?;
+
+  generate_image(bytes, prompt, &app_data_root)
+    .await
+    .map_err(|err| {
+      error!("error: {:?}", err);
+      "there was an error".to_string()
+    })?;
+
+  Ok("success".to_string())
+}
+
+pub async fn generate_image(file_bytes: Vec<u8>, prompt: &str, app_data_root: &AppDataRoot) -> AnyhowResult<()> {
+  let sora_credentials = get_credentials(app_data_root)?;
+
+  let filename = "image.png".to_string();
+
+  let response= sora_media_upload_from_bytes(file_bytes, filename, &sora_credentials)
+      .await
+      .map_err(|err| {
+        error!("Failed to upload image to Sora: {:?}", err);
+        err
+      })?;
+
+  let sora_media_tokens = vec![response.id];
+
+  let response = sora_image_gen_remix(SoraImageGenRemixRequest {
+    prompt: prompt.to_string(),
+    num_images: NumImages::One,
+    image_size: ImageSize::Square,
+    sora_media_tokens: sora_media_tokens.clone(),
+    credentials: &sora_credentials,
+  }).await
+      .map_err(|err| {
+        error!("Failed to call Sora image generation: {:?}", err);
+        err
+      })?;
+
+  println!(">> TASK ID: {:?} ", response.task_id);
+
+  Ok(())
+}
+
+fn get_credentials(app_data_root: &AppDataRoot) -> AnyhowResult<SoraCredentials> {
+  let cookie_file = app_data_root.get_sora_cookie_file_path();
+
+  let bearer = read_to_string("/Users/bt/dev/storyteller/storyteller-rust/test_data/temp/bearer.txt")?
+      .trim()
+      .to_string();
+
+  let cookie= read_to_string(cookie_file)?
+      .trim()
+      .to_string();
+
+  let sentinel = read_to_string("/Users/bt/dev/storyteller/storyteller-rust/test_data/temp/sentinel.txt")?
+      .trim()
+      .to_string();
+
+  Ok(SoraCredentials {
+    bearer_token: bearer,
+    cookie: cookie,
+    sentinel: Some(sentinel),
+  })
+}

crates/desktop/tauri-4o/src/commands/mod.rs

@@ -0,0 +1,3 @@
+pub mod flip_image;
+pub mod image_generation_command;
+pub mod open_login_command;

crates/desktop/tauri-4o/src/commands/open_login_command.rs

@@ -0,0 +1,72 @@
+use std::time::Duration;
+use log::{error, info};
+use once_cell::sync::Lazy;
+use tauri::{AppHandle, Manager, Url, WebviewUrl, WebviewWindowBuilder, WindowBuilder};
+use tokio::time::sleep_until;
+use errors::AnyhowResult;
+use crate::threads::login_thread::{LOGIN_WINDOW_NAME, SORA_LOGIN_URL};
+use crate::utils::clear_all_webview_cookies::clear_all_webview_cookies;
+
+pub const START_URL_STR: &str = "https://storyteller.ai/";
+
+pub static START_URL: Lazy<Url> = Lazy::new(|| {
+  Url::parse(START_URL_STR).expect("URL should parse")
+});
+
+pub const SORA_HOMEPAGE_URL_STR: &str = "https://sora.com/";
+
+pub static SORA_HOMEPAGE_URL : Lazy<Url> = Lazy::new(|| {
+  Url::parse(SORA_HOMEPAGE_URL_STR).expect("URL should parse")
+});
+
+#[tauri::command]
+pub async fn open_login_command(
+  app: AppHandle,
+) -> Result<String, String> {
+  info!("open_login_command called");
+
+  do_open_login(&app)
+    .await
+    .map_err(|err| {
+      error!("Error opening login: {:?}", err);
+      format!("Error opening login: {:?}", err)
+    })?;
+
+  Ok("result".to_string())
+}
+
+async fn do_open_login(app: &AppHandle) -> AnyhowResult<()> {
+  info!("Building login window...");
+
+  let url = WebviewUrl::External(START_URL.clone());
+
+  let window = WebviewWindowBuilder::new(app, LOGIN_WINDOW_NAME, url)
+      //.user_agent(openai_sora_client::credentials::USER_AGENT)
+      .always_on_top(true)
+      .title("Login to OpenAI")
+      .center()
+      .resizable(true)
+      .visible(true)
+      .closable(true)
+      .focused(true)
+      .devtools(true)
+      .build()?;
+
+  for (name, webview) in window.webviews() {
+    if name != LOGIN_WINDOW_NAME {
+      continue;
+    }
+
+    clear_all_webview_cookies(&webview)?;
+
+    // NB: We're starting to get Cloudflare protection screens. Let's try to avoid.
+    webview.navigate(SORA_HOMEPAGE_URL.clone())?;
+    tokio::time::sleep(Duration::from_millis(100)).await;
+
+    webview.navigate(SORA_LOGIN_URL.clone())?;
+    break;
+  }
+
+  info!("Done.");
+  Ok(())
+}

crates/desktop/tauri-4o/src/events/notification_event.rs

@@ -0,0 +1,45 @@
+use serde::Serialize;
+
+#[derive(Clone, Copy, Serialize)]
+#[serde(rename_all = "snake_case")]
+pub enum NotificationModelType {
+  ImageSegmentation,
+  TextEncoder,
+  TextTokenizer,
+  Unet,
+  Vae,
+  Json,
+  ClipEncoder,
+  FluxModel,
+}
+
+#[derive(Clone, Serialize)]
+#[serde(rename_all = "snake_case")]
+pub enum NotificationEvent<'a> {
+  ModelDownloadStarted {
+    model_name: &'a str, 
+    model_type: NotificationModelType,
+  },
+  ModelDownloadProgress {
+    model_name: &'a str,
+    model_type: NotificationModelType,
+    currently_downloaded_bytes: usize,
+    total_file_bytes: usize,
+  },
+  ModelDownloadComplete {
+    model_name: &'a str,
+    model_type: NotificationModelType,
+  },
+  ModelLoading {
+    model_name: String,
+    model_type: NotificationModelType,
+  },
+  ModelInference {
+    model_name: String,
+    model_type: NotificationModelType,
+  },
+  ModelInferenceComplete {
+    model_name: String,
+    model_type: NotificationModelType,
+  },
+}

crates/desktop/tauri-4o/src/lib.rs

@@ -0,0 +1,65 @@
+pub mod commands;
+pub mod events;
+pub mod state;
+pub mod threads;
+pub mod transfer;
+pub mod utils;
+
+use crate::commands::flip_image::flip_image;
+use crate::commands::image_generation_command::image_generation_command;
+use crate::commands::open_login_command::open_login_command;
+use crate::state::app_config::AppConfig;
+use crate::threads::login_thread::login_thread;
+
+use tauri_plugin_log::Target;
+use tauri_plugin_log::TargetKind;
+
+#[cfg_attr(mobile, tauri::mobile_entry_point)]
+pub fn run() {
+  // NB: Tauri wants to install the logger itself, so we can't rely on the logger crate 
+  // until the tauri runtime begins.
+  println!("Loading model config...");
+
+  let config = AppConfig::init()
+    .expect("config should load");
+
+  let app_data_root = config.app_data_root.clone();
+  let app_data_root2 = config.app_data_root.clone();
+
+  println!("Initializing backend runtime...");
+
+  tauri::Builder::default()
+    .plugin(tauri_plugin_log::Builder::new()
+      .level(log::LevelFilter::Info)
+      .targets(vec![
+        Target::new(TargetKind::Stdout),
+        Target::new(TargetKind::LogDir { file_name: Some(app_data_root.log_file_name_str().to_string()) }),
+      ])
+      .build())
+    .setup(|app| {
+      // TODO(bt): This is broken on windows
+      // log_environment_details();
+
+      //if cfg!(debug_assertions) {
+      //  app.handle().plugin(
+      //    tauri_plugin_log::Builder::default()
+      //      .level(log::LevelFilter::Info)
+      //      .build(),
+      //  )?;
+      //}
+      let app = app.handle().clone();
+
+      tauri::async_runtime::spawn(login_thread(app, app_data_root2));
+
+      Ok(())
+    })
+    .manage(config)
+    .manage(app_data_root)
+    .invoke_handler(tauri::generate_handler![
+      flip_image,
+      image_generation_command,
+      open_login_command,
+    ])
+    .run(tauri::generate_context!("tauri.conf.json"))
+    .expect("error while running tauri application");
+}

crates/desktop/tauri-4o/src/main.rs

@@ -0,0 +1,6 @@
+// Prevents additional console window on Windows in release, DO NOT REMOVE!!
+#![cfg_attr(not(debug_assertions), windows_subsystem = "windows")]
+
+fn main() {
+  app_lib::run();
+}

crates/desktop/tauri-4o/src/state/app_config.rs

@@ -0,0 +1,58 @@
+use crate::state::app_dir::AppDataRoot;
+use crate::state::os_platform::OsPlatform;
+use crate::state::yaml_config::YamlConfig;
+
+const DEFAULT_SD_IMAGE_WIDTH: usize = 1024;
+const DEFAULT_SD_IMAGE_HEIGHT: usize = 1024;
+
+
+/// A central place to configure app-wide details.
+pub struct AppConfig {
+  pub image_height: usize,
+  pub image_width: usize,
+  pub scheduler_steps: usize,
+  pub scheduler_samples: usize,
+  
+  pub seed: Option<u64>,
+
+  pub cfg_scale: Option<f64>,
+  
+  /// Root location for application data.
+  pub app_data_root: AppDataRoot,
+}
+
+impl AppConfig {
+  
+  pub fn init() -> anyhow::Result<Self> {
+    let yaml_configs = YamlConfig::load_from_config_file_recursive();
+    
+    println!("Configs: {:?}", yaml_configs);
+    
+    let os_platform = OsPlatform::get();
+
+    let default_data_root = match os_platform {
+      OsPlatform::Linux => yaml_configs.linux_default_data_path.as_deref(),
+      OsPlatform::MacOs => yaml_configs.mac_default_data_path.as_deref(),
+      OsPlatform::Windows => yaml_configs.windows_default_data_path.as_deref(),
+    };
+    
+    println!("Possible user-defined data root: {:?}", default_data_root);
+
+    let app_data_root = match default_data_root {
+      None => AppDataRoot::create_default()?,
+      Some(path) => AppDataRoot::create_existing(path)?,
+    };
+
+    println!("App data root: {:?}", app_data_root.path());
+
+    Ok(Self {
+      image_height: yaml_configs.image_height.unwrap_or(DEFAULT_SD_IMAGE_HEIGHT),
+      image_width: yaml_configs.image_width.unwrap_or(DEFAULT_SD_IMAGE_WIDTH),
+      scheduler_steps: yaml_configs.scheduler_steps.unwrap_or(4),
+      scheduler_samples: yaml_configs.scheduler_samples.unwrap_or(15),
+      seed: yaml_configs.seed,
+      cfg_scale: yaml_configs.cfg_scale,
+      app_data_root,
+    })
+  }
+}

crates/desktop/tauri-4o/src/state/app_dir.rs

@@ -0,0 +1,219 @@
+use crate::state::expanduser::expanduser;
+use crate::state::os_platform::OsPlatform;
+use anyhow::anyhow;
+use directories::UserDirs;
+use std::path::{Path, PathBuf};
+use chrono::{DateTime, Local};
+use tempdir::TempDir;
+use tempfile::{Builder, NamedTempFile};
+
+const DEFAULT_DATA_DIR : &str = "artcraft";
+const ASSETS_SUBDIRECTORY : &str = "assets";
+const WEIGHTS_SUBDIRECTORY : &str = "weights";
+
+const TEMPORARY_SUBDIRECTORY : &str = "temp";
+
+/// Note: Tauri appends ".log" to the end of the filename.
+const LOG_FILE_NAME : &str = "application_debug";
+
+/// The path to the application data directory, which includes "asset" and "weights" data.
+#[derive(Clone)]
+pub struct AppDataRoot {
+  path: PathBuf,
+  assets_dir: AppAssetsDir,
+  weights_dir: AppWeightsDir,
+  temp_dir: TemporaryDir,
+  log_file_name: PathBuf,
+  log_file_name_string: String,
+}
+
+#[derive(Clone)]
+pub struct AppAssetsDir {
+  path: PathBuf,
+}
+
+#[derive(Clone)]
+pub struct AppWeightsDir {
+  path: PathBuf,
+}
+
+#[derive(Clone)]
+pub struct TemporaryDir {
+  path: PathBuf,
+}
+
+impl AppDataRoot {
+  pub fn create_default() -> anyhow::Result<Self> {
+    let directory = get_default_data_dir()?;
+    Self::create_existing(directory)
+  }
+
+  pub fn create_existing<P: AsRef<Path>>(dir: P) -> anyhow::Result<Self> {
+    let mut dir = dir.as_ref().to_path_buf();
+    
+    match OsPlatform::get() {
+      OsPlatform::Linux | OsPlatform::MacOs => {
+        if let Some(d) = dir.as_os_str().to_str() {
+          dir = expanduser(d)?;
+        }
+      },
+      OsPlatform::Windows => {}
+    }
+    
+    if !dir.is_dir() {
+      println!("Creating directory {:?}", dir);
+      std::fs::create_dir_all(&dir)?;
+    }
+
+    match dir.canonicalize() {
+      Ok(d) => dir = d,
+      Err(err) => {
+        println!("Error canonicalizing {:?}: {}", dir, err);
+      }
+    }
+    
+    let assets_dir = AppAssetsDir::create_existing(dir.join(ASSETS_SUBDIRECTORY))?;
+    let weights_dir = AppWeightsDir::create_existing(dir.join(WEIGHTS_SUBDIRECTORY))?;
+    let temp_dir = TemporaryDir::create_existing(dir.join(TEMPORARY_SUBDIRECTORY))?;
+    let log_file_name = dir.join(LOG_FILE_NAME);
+    let log_file_name_string = log_file_name
+        .to_str()
+        .ok_or(anyhow!("couldn't convert log path to str"))?
+        .to_string();
+
+    Ok(Self {
+      path: dir,
+      assets_dir,
+      weights_dir,
+      temp_dir,
+      log_file_name,
+      log_file_name_string,
+    })
+  }
+  
+  pub fn assets_dir(&self) -> &AppAssetsDir {
+    &self.assets_dir
+  }
+  
+  pub fn weights_dir(&self) -> &AppWeightsDir {
+    &self.weights_dir
+  }
+
+  pub fn temp_dir(&self) -> &TemporaryDir {
+    &self.temp_dir
+  }
+
+  pub fn path(&self) -> &Path {
+    &self.path
+  }
+
+  pub fn log_file_name(&self) -> &Path {
+    &self.log_file_name
+  }
+
+  pub fn log_file_name_str(&self) -> &str {
+    &self.log_file_name_string
+  }
+
+  pub fn get_sora_cookie_file_path(&self) -> PathBuf {
+    self.path.join("sora_cookies.txt")
+  }
+}
+
+impl AppAssetsDir {
+  pub fn create_existing<P: AsRef<Path>>(dir: P) -> anyhow::Result<Self> {
+    let mut dir = dir.as_ref().to_path_buf();
+    match dir.canonicalize() {
+      Ok(d) => dir = d,
+      Err(err) => {
+        println!("Error canonicalizing {:?}: {}", dir, err);
+      }
+    }
+    if !dir.exists() {
+      println!("Creating directory {:?}", dir);
+      std::fs::create_dir(&dir)?;
+    }
+    Ok(Self {
+      path: dir,
+    })
+  }
+
+  pub fn path(&self) -> &Path {
+    &self.path
+  }
+  
+  /// Return the current date asset directory
+  /// If it doesn't exist, create it.
+  pub fn make_or_get_current_date_dir(&self) -> anyhow::Result<PathBuf> {
+    let format: DateTime<Local> = Local::now();
+    let date_directory_name = format.format("%Y-%d-%m").to_string();
+    let full_path = self.path.join(date_directory_name);
+    if !full_path.exists() {
+      std::fs::create_dir(&full_path)?;
+    }
+    Ok(full_path)
+  }
+}
+
+impl AppWeightsDir {
+  pub fn create_existing<P: AsRef<Path>>(dir: P) -> anyhow::Result<Self> {
+    let mut dir = dir.as_ref().to_path_buf();
+    match dir.canonicalize() {
+      Ok(d) => dir = d,
+      Err(err) => {
+        println!("Error canonicalizing {:?}: {}", dir, err);
+      }
+    }
+    if !dir.exists() {
+      println!("Creating directory {:?}", dir);
+      std::fs::create_dir(&dir)?;
+    }
+    Ok(Self {
+      path: dir,
+    })
+  }
+
+  pub fn path(&self) -> &Path {
+    &self.path
+  }
+}
+
+impl TemporaryDir {
+  pub fn create_existing<P: AsRef<Path>>(dir: P) -> anyhow::Result<Self> {
+    let mut dir = dir.as_ref().to_path_buf();
+    match dir.canonicalize() {
+      Ok(d) => dir = d,
+      Err(err) => {
+        println!("Error canonicalizing {:?}: {}", dir, err);
+      }
+    }
+    if !dir.exists() {
+      println!("Creating directory {:?}", dir);
+      std::fs::create_dir(&dir)?;
+    }
+    Ok(Self {
+      path: dir,
+    })
+  }
+
+  pub fn with_prefix(&self, prefix: &str) -> anyhow::Result<TempDir> {
+    let tempdir = TempDir::new_in(&self.path, prefix)?;
+    Ok(tempdir)
+  }
+
+  pub fn new_named_temp_file(&self) -> anyhow::Result<NamedTempFile> {
+    let tempfile = Builder::new()
+      .prefix("temp_")
+      .suffix(".bin")
+      .tempfile_in(&self.path)?;
+    Ok(tempfile)
+  }
+}
+
+// eg. /home/bob/artcraft, /Users/bob/artcraft, or C:\Users\Alice\artcraft
+fn get_default_data_dir() -> anyhow::Result<PathBuf> {
+  Ok(UserDirs::new()
+      .ok_or_else(|| anyhow!("could not determine user home directory"))?
+      .home_dir()
+      .join(DEFAULT_DATA_DIR))
+}

crates/desktop/tauri-4o/src/state/expanduser.rs

@@ -0,0 +1,73 @@
+use once_cell::sync::Lazy;
+use std::io;
+use std::path::{PathBuf, MAIN_SEPARATOR};
+
+static PREFIX: Lazy<String> = Lazy::new(|| format!("~{}", MAIN_SEPARATOR));
+
+/// The "expanduser" crate doesn't compile on Windows, so we replace its functionality slightly
+pub fn expanduser<P: AsRef<str>>(path: P) -> io::Result<PathBuf> {
+
+   Ok(match path.as_ref() {
+       // matches an exact "~"
+       s if s == "~" => {
+           home_dir()?
+       },
+       // matches paths that start with `~/`
+       s if s.starts_with(&*PREFIX) => {
+           let home = home_dir()?;
+           home.join(&s[2..])
+       },
+       // // matches paths that start with `~` but not `~/`, might be a `~username/` path
+       // s if s.starts_with("~") => {
+       //     let mut parts = s[1..].splitn(2, MAIN_SEPARATOR);
+       //     let user = parts.next()
+       //         .ok_or_else(|| io::Error::new(io::ErrorKind::Other, "malformed path"))?;
+       //     let user = Passwd::from_name(&user)
+       //         .map_err(|_| io::Error::new(io::ErrorKind::Other, "error searching for user"))?
+       //         .ok_or_else(|| io::Error::new(io::ErrorKind::Other, format!("user '{}', does not exist", &user)))?;
+       //     if let Some(ref path) = parts.next() {
+       //         PathBuf::from(user.dir).join(&path)
+       //     } else {
+       //         PathBuf::from(user.dir)
+       //     }
+       // },
+       // nothing to expand, just make a PathBuf
+       s => PathBuf::from(s)
+   })
+}
+
+pub fn home_dir() -> io::Result<PathBuf> {
+    dirs::home_dir()
+        .ok_or_else(|| io::Error::new(io::ErrorKind::Other, "no home directory is set"))
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::state::expanduser::expanduser;
+    use std::env;
+    use std::path::PathBuf;
+
+    #[test]
+    #[serial_test::serial]
+    fn test_success() {
+        let old_home = env::var("HOME").expect("no home dir set");
+        let new_home = "/home/foo";
+        env::set_var("HOME", new_home);
+        let path = expanduser("~/path/to/directory");
+        env::set_var("HOME", old_home);
+        assert_eq!(path.expect("io error"), PathBuf::from("/home/foo/path/to/directory"));
+    }
+
+    #[test]
+    #[serial_test::serial]
+    fn test_only_tilde() {
+        let old_home = env::var("HOME").expect("no home dir set");
+        let new_home = "/home/foo";
+        env::set_var("HOME", new_home);
+        let pathstr = "~";
+        let path = expanduser(pathstr);
+        env::set_var("HOME", old_home);
+        assert_eq!(path.expect("io error"), PathBuf::from("/home/foo"));
+    }
+
+}

crates/desktop/tauri-4o/src/state/mod.rs

@@ -1,4 +1,5 @@
 pub mod app_config;
 pub mod app_dir;
 pub mod os_platform;
-pub mod yaml_config;
+pub mod yaml_config;
+pub mod expanduser;

crates/desktop/tauri-4o/src/threads/login_thread.rs

@@ -0,0 +1,103 @@
+use std::fs::OpenOptions;
+use std::io::Write;
+use anyhow::anyhow;
+use log::{error, info};
+use once_cell::sync::Lazy;
+use reqwest::Url;
+use tauri::{AppHandle, Manager, Webview};
+use errors::AnyhowResult;
+use crate::state::app_dir::AppDataRoot;
+use crate::utils::sora_webview_cookies::get_all_sora_cookies_as_string;
+
+pub const LOGIN_WINDOW_NAME: &str = "login_window";
+
+pub const SORA_LOGIN_URL_STR: &str = "https://sora.com/auth/login?callback_path=/";
+
+pub static SORA_LOGIN_URL: Lazy<Url> = Lazy::new(|| {
+  Url::parse(SORA_LOGIN_URL_STR).expect("URL should parse")
+});
+
+pub const SORA_ROOT_URL_STR: &str = "https://sora.com/";
+
+pub static SORA_ROOT_URL: Lazy<Url> = Lazy::new(|| {
+  Url::parse(SORA_ROOT_URL_STR).expect("URL should parse")
+});
+
+pub async fn login_thread(app: AppHandle, app_data_root: AppDataRoot) -> ! {
+  loop {
+    for (window_name, webview) in app.webviews() {
+      if window_name == LOGIN_WINDOW_NAME {
+        let result = check_login_window(&webview, &app_data_root).await;
+        if let Err(err) = result {
+          error!("Error checking login window: {:?}", err);
+        }
+        break;
+      }
+    }
+    tokio::time::sleep(std::time::Duration::from_millis(2_000)).await;
+  }
+}
+
+async fn check_login_window(webview: &Webview, app_data_root: &AppDataRoot) -> AnyhowResult<()> {
+  clear_browsing_data_on_test_domain(webview)?;
+  //keep_on_task(webview)?;
+  extract_cookies_to_file(webview, app_data_root)?;
+  // TODO: Other maintenance tasks ...
+  Ok(())
+}
+
+fn keep_on_task(webview: &Webview) -> AnyhowResult<()> {
+  let url = webview.url()?;
+  let hostname= url.host()
+      .ok_or(anyhow!("no host in url"))?
+      .to_string();
+  match hostname.as_str() {
+    "auth.openai.com" => {},
+    "openai.com" => {},
+    "sora.com" => {},
+    // Third party SSO
+    "accounts.google.com" => {},
+    "accounts.youtube.com" => {},
+    "login.live.com" => {},
+    "appleid.apple.com" => {},
+    _ => {
+      info!("Non login hostname: {}", hostname);
+      webview.navigate(SORA_LOGIN_URL.clone())?;
+    }
+  }
+  Ok(())
+}
+
+/// This is just so we have a way to clear browsing data.
+fn clear_browsing_data_on_test_domain(webview: &Webview) -> AnyhowResult<()> {
+  let url = webview.url()?;
+  let hostname= url.host()
+      .ok_or(anyhow!("no host in url"))?
+      .to_string();
+  match hostname.as_str() {
+    "storyteller.ai" => {
+      info!("Clearing all browsing data...");
+      webview.clear_all_browsing_data()?;
+    }
+    _ => {}
+  }
+  Ok(())
+}
+
+// TODO(bt,2025-04-07): Heuristic to detect when logged in. Only write when logged in.
+fn extract_cookies_to_file(webview: &Webview, app_data_root: &AppDataRoot) -> AnyhowResult<()> {
+  let cookies = get_all_sora_cookies_as_string(webview)?;
+
+  let cookie_file = app_data_root.get_sora_cookie_file_path();
+
+  let mut file = OpenOptions::new()
+      .create(true)
+      .write(true)
+      .truncate(true)
+      .open(cookie_file)?;
+
+  file.write_all(cookies.as_bytes())?;
+  file.flush()?;
+
+  Ok(())
+}

crates/desktop/tauri-4o/src/threads/mod.rs

@@ -0,0 +1 @@
+pub mod login_thread;

crates/desktop/tauri-4o/src/transfer/download.rs

@@ -0,0 +1,241 @@
+//! Implemented from https://github.com/huggingface/hf_transfer/blob/main/src/lib.rs
+
+use crate::transfer::error::Error;
+use crate::transfer::util::{exponential_backoff, BASE_WAIT_TIME, MAX_WAIT_TIME};
+use futures::stream::FuturesUnordered;
+use futures::StreamExt;
+use reqwest::header::{
+  HeaderMap, HeaderName, HeaderValue, AUTHORIZATION, CONTENT_RANGE,
+  RANGE,
+};
+use reqwest::Url;
+use std::collections::HashMap;
+use std::fmt::Display;
+use std::fs::remove_file;
+use std::io::SeekFrom;
+use std::path::Path;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::mpsc::Sender;
+use std::time::Duration;
+use log::info;
+use tempfile::NamedTempFile;
+use tokio::fs::OpenOptions;
+use tokio::io::AsyncWriteExt;
+use tokio::io::AsyncSeekExt;
+use tokio::sync::Semaphore;
+use tokio::time::sleep;
+
+pub struct ProgressUpdate {
+  pub complete: usize,
+  pub total_length: usize,
+}
+
+/// max_files: Number of open file handles, which determines the maximum number of parallel downloads
+/// parallel_failures:  Number of maximum failures of different chunks in parallel (cannot exceed max_files)
+/// max_retries: Number of maximum attempts per chunk. (Retries are exponentially backed off + jitter)
+///
+/// The number of threads can be tuned by the environment variable `TOKIO_WORKER_THREADS` as documented in
+/// https://docs.rs/tokio/latest/tokio/runtime/struct.Builder.html#method.worker_threads
+#[allow(clippy::too_many_arguments)]
+pub async fn download_async<P: AsRef<Path>>(
+  url: String,
+  filename: NamedTempFile,
+  final_filename: P,
+  max_files: usize,
+  chunk_size: usize,
+  parallel_failures: usize,
+  max_retries: usize,
+  input_headers: Option<HashMap<String, String>>,
+  maybe_progress_sender: Option<Sender<ProgressUpdate>>,
+) -> Result<(), Error> {
+  info!(">>> Downloading: {}", filename.as_ref().display());
+
+  let client = reqwest::Client::builder()
+    // https://github.com/hyperium/hyper/issues/2136#issuecomment-589488526
+    .http2_keep_alive_timeout(Duration::from_secs(15))
+    .build()?;
+
+  let mut headers = HeaderMap::new();
+  let mut auth_token = None;
+  if let Some(input_headers) = input_headers {
+    headers.reserve(input_headers.len());
+    for (k, v) in input_headers {
+      let name: HeaderName = k
+        .try_into()
+        .map_err(|err| Error::new_err(format!("Invalid header: {err}")))?;
+      let value: HeaderValue = AsRef::<str>::as_ref(&v)
+        .try_into()
+        .map_err(|err| Error::new_err(format!("Invalid header value: {err}")))?;
+      if name == AUTHORIZATION {
+        auth_token = Some(value);
+      } else {
+        headers.insert(name, value);
+      }
+    }
+  };
+
+  let response = if let Some(token) = auth_token.as_ref() {
+    client.get(&url).header(AUTHORIZATION, token)
+  } else {
+    client.get(&url)
+  }
+    .headers(headers.clone())
+    .header(RANGE, "bytes=0-0")
+    .send()
+    .await
+    .map_err(|err| Error::new_err(format!("Error while downloading: {err}")))?
+    .error_for_status()
+    .map_err(|err| Error::new_err(err.to_string()))?;
+
+  // Only call the final redirect URL to avoid overloading the Hub with requests and also
+  // altering the download count
+  let redirected_url = response.url();
+  if Url::parse(&url)
+    .map_err(|err| Error::new_err(format!("failed to parse url: {err}")))?
+    .host()
+    == redirected_url.host()
+  {
+    if let Some(token) = auth_token {
+      headers.insert(AUTHORIZATION, token);
+    }
+  }
+
+  let content_range = response
+    .headers()
+    .get(CONTENT_RANGE)
+    .ok_or(Error::new_err("No content length".to_string()))?
+    .to_str()
+    .map_err(|err| Error::new_err(format!("Error while downloading: {err}")))?;
+
+  let size: Vec<&str> = content_range.split('/').collect();
+  // Content-Range: bytes 0-0/702517648
+  // https://developer.mozilla.org/en-US/docs/Web/HTTP/Headers/Content-Range
+  let length: usize = size
+    .last()
+    .ok_or(Error::new_err(
+      "Error while downloading: No size was detected".to_string(),
+    ))?
+    .parse()
+    .map_err(|err| Error::new_err(format!("Error while downloading: {err}")))?;
+
+
+  info!("File {:?} is {:?} length", filename.as_ref(), length);
+
+  let mut handles = FuturesUnordered::new();
+  let semaphore = Arc::new(Semaphore::new(max_files));
+  let parallel_failures_semaphore = Arc::new(Semaphore::new(parallel_failures));
+  let max_chunk_complete = Arc::new(AtomicUsize::new(0));
+
+  for start in (0..length).step_by(chunk_size) {
+    let url = redirected_url.to_string();
+    let filename = filename.as_ref().to_owned();
+    let client = client.clone();
+    let headers = headers.clone();
+
+    let stop = std::cmp::min(start + chunk_size - 1, length);
+    let semaphore = semaphore.clone();
+    let parallel_failures_semaphore = parallel_failures_semaphore.clone();
+    let maybe_progress = maybe_progress_sender.clone();
+    let max_chunk_complete = max_chunk_complete.clone();
+    handles.push(tokio::spawn(async move {
+      let permit = semaphore
+        .acquire_owned()
+        .await
+        .map_err(|err| Error::new_err(format!("Error while downloading: {err}")))?;
+      let mut chunk = download_chunk(&client, &url, &filename, start, stop, headers.clone()).await;
+      let mut i = 0;
+      if parallel_failures > 0 {
+        while let Err(dlerr) = chunk {
+          if i >= max_retries {
+            return Err(Error::new_err(format!(
+              "Failed after too many retries ({max_retries}): {dlerr}"
+            )));
+          }
+          let parallel_failure_permit = parallel_failures_semaphore.clone().try_acquire_owned().map_err(|err| {
+            Error::new_err(format!(
+              "Failed too many failures in parallel ({parallel_failures}): {dlerr} ({err})"
+            ))
+          })?;
+
+          let wait_time = exponential_backoff(BASE_WAIT_TIME, i, MAX_WAIT_TIME);
+          sleep(Duration::from_millis(wait_time as u64)).await;
+
+          chunk = download_chunk(&client, &url, &filename, start, stop, headers.clone()).await;
+          i += 1;
+          drop(parallel_failure_permit);
+          
+          if let Some(sender) = maybe_progress.as_ref() {
+            // NB: Calling fetch_max twice. First call will return the lower value.
+            let _old = max_chunk_complete.fetch_max(stop, Ordering::SeqCst);
+            let max_chunk = max_chunk_complete.fetch_max(stop, Ordering::SeqCst);
+            let progress_update = ProgressUpdate { 
+              complete: max_chunk,
+              total_length: length,
+            };
+            if let Err(_err) = sender.send(progress_update) {
+              // Fail silently, not essential to issue progress updates
+            }
+          }
+        }
+      }
+      drop(permit);
+      chunk.map_err(|e| Error::new_err(format!("Downloading error {e}"))).and(Ok(stop - start))
+    }));
+  }
+
+  // Output the chained result
+  while let Some(result) = handles.next().await {
+    match result {
+      Ok(Ok(size)) => {
+        //if let Some(ref callback) = callback {
+        //  callback.call((size,), None)?;
+        //}
+      }
+      Ok(Err(py_err)) => {
+        return Err(py_err);
+      }
+      Err(err) => {
+        return Err(Error::new_err(format!(
+          "Error while downloading: {err}"
+        )));
+      }
+    }
+  }
+
+  let file = filename.persist(final_filename)
+    .map_err(|e| Error::new_err(format!("Could not rename to final filename: {:?}", e)))?;
+
+  file.sync_all()?;
+
+  Ok(())
+}
+
+async fn download_chunk<P: AsRef<Path>>(
+  client: &reqwest::Client,
+  url: &str,
+  filename: P,
+  start: usize,
+  stop: usize,
+  headers: HeaderMap,
+) -> Result<(), Error> {
+  // Process each socket concurrently.
+  let range = format!("bytes={start}-{stop}");
+  let mut file = OpenOptions::new()
+    .write(true)
+    .truncate(false)
+    .create(true)
+    .open(filename)
+    .await?;
+  file.seek(SeekFrom::Start(start as u64)).await?;
+  let response = client
+    .get(url)
+    .headers(headers)
+    .header(RANGE, range)
+    .send()
+    .await?
+    .error_for_status()?;
+  let content = response.bytes().await?;
+  file.write_all(&content).await?;
+  Ok(())
+}

crates/desktop/tauri-4o/src/transfer/error.rs

@@ -0,0 +1,49 @@
+//! Implemented from https://github.com/huggingface/hf_transfer/blob/main/src/lib.rs
+
+use std::fmt::Display;
+use reqwest::header::ToStrError;
+
+#[derive(Debug)]
+pub enum Error {
+  Io(std::io::Error),
+  Request(reqwest::Error),
+  ToStrError(ToStrError),
+  Misc(String),
+}
+
+impl From<std::io::Error> for Error {
+  fn from(value: std::io::Error) -> Self {
+    Self::Io(value)
+  }
+}
+
+impl From<reqwest::Error> for Error {
+  fn from(value: reqwest::Error) -> Self {
+    Self::Request(value)
+  }
+}
+
+impl From<ToStrError> for Error {
+  fn from(value: ToStrError) -> Self {
+    Self::ToStrError(value)
+  }
+}
+
+impl Display for Error {
+  fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+    match self {
+      Self::Io(io) => write!(f, "Io: {io}"),
+      Self::Request(req) => write!(f, "Request: {req}"),
+      Self::ToStrError(req) => write!(f, "Response non ascii: {req}"),
+      Self::Misc(err) => write!(f, "Misc error: {err}"),
+    }
+  }
+}
+
+impl std::error::Error for Error {}
+
+impl Error {
+  pub fn new_err(reason: String) -> Self {
+    Self::Misc(reason)
+  }
+}

crates/desktop/tauri-4o/src/transfer/mod.rs

@@ -0,0 +1,4 @@
+pub mod download;
+pub mod upload;
+pub mod util;
+pub mod error;

crates/desktop/tauri-4o/src/transfer/upload.rs

@@ -0,0 +1,189 @@
+//! Implemented from https://github.com/huggingface/hf_transfer/blob/main/src/lib.rs
+
+use crate::transfer::error::Error;
+use crate::transfer::util::{exponential_backoff, BASE_WAIT_TIME, MAX_WAIT_TIME};
+use futures::stream::FuturesUnordered;
+use futures::StreamExt;
+use reqwest::header::CONTENT_LENGTH;
+use std::collections::HashMap;
+use std::fmt::Display;
+use std::io::SeekFrom;
+use std::sync::Arc;
+use std::time::Duration;
+use tokio::fs::OpenOptions;
+use tokio::io::{AsyncReadExt, AsyncSeekExt};
+use tokio::sync::Semaphore;
+use tokio::time::sleep;
+use tokio_util::codec::{BytesCodec, FramedRead};
+
+/// parts_urls: Dictionary consisting of part numbers as keys and the associated url as values
+/// completion_url: The url that should be called when the upload is finished
+/// max_files: Number of open file handles, which determines the maximum number of parallel uploads
+/// parallel_failures:  Number of maximum failures of different chunks in parallel (cannot exceed max_files)
+/// max_retries: Number of maximum attempts per chunk. (Retries are exponentially backed off + jitter)
+///
+/// The number of threads can be tuned by the environment variable `TOKIO_WORKER_THREADS` as documented in
+/// https://docs.rs/tokio/latest/tokio/runtime/struct.Builder.html#method.worker_threads
+///
+/// See https://docs.aws.amazon.com/AmazonS3/latest/userguide/mpuoverview.html for more information
+/// on the multipart upload
+#[allow(clippy::too_many_arguments)]
+fn multipart_upload(
+  file_path: String,
+  parts_urls: Vec<String>,
+  chunk_size: u64,
+  max_files: usize,
+  parallel_failures: usize,
+  max_retries: usize,
+) -> Result<Vec<HashMap<String, String>>, Error> {
+  if parallel_failures > max_files {
+    return Err(Error::new_err(
+      "Error parallel_failures cannot be > max_files".to_string(),
+    ));
+  }
+  if (parallel_failures == 0) != (max_retries == 0) {
+    return Err(Error::new_err(
+      "For retry mechanism you need to set both `parallel_failures` and `max_retries`"
+        .to_string(),
+    ));
+  }
+
+  tokio::runtime::Builder::new_multi_thread()
+    .enable_all()
+    .build()?
+    .block_on(async {
+      upload_async(
+        file_path,
+        parts_urls,
+        chunk_size,
+        max_files,
+        parallel_failures,
+        max_retries,
+      )
+        .await
+    })
+}
+
+
+
+#[allow(clippy::too_many_arguments)]
+async fn upload_async(
+  file_path: String,
+  parts_urls: Vec<String>,
+  chunk_size: u64,
+  max_files: usize,
+  parallel_failures: usize,
+  max_retries: usize,
+) -> Result<Vec<HashMap<String, String>>, Error> {
+  let client = reqwest::Client::new();
+
+  let mut handles = FuturesUnordered::new();
+  let semaphore = Arc::new(Semaphore::new(max_files));
+  let parallel_failures_semaphore = Arc::new(Semaphore::new(parallel_failures));
+
+  for (part_number, part_url) in parts_urls.iter().enumerate() {
+    let url = part_url.to_string();
+    let path = file_path.to_owned();
+    let client = client.clone();
+
+    let start = (part_number as u64) * chunk_size;
+    let semaphore = semaphore.clone();
+    let parallel_failures_semaphore = parallel_failures_semaphore.clone();
+    handles.push(tokio::spawn(async move {
+      let permit = semaphore
+        .clone()
+        .acquire_owned()
+        .await
+        .map_err(|err| Error::new_err(format!("Error acquiring semaphore: {err}")))?;
+      let mut chunk = upload_chunk(&client, &url, &path, start, chunk_size).await;
+      let mut i = 0;
+      if parallel_failures > 0 {
+        while let Err(ul_err) = chunk {
+          if i >= max_retries {
+            return Err(Error::new_err(format!(
+              "Failed after too many retries ({max_retries}): {ul_err}"
+            )));
+          }
+
+          let parallel_failure_permit = parallel_failures_semaphore.clone().try_acquire_owned().map_err(|err| {
+            Error::new_err(format!(
+              "Failed too many failures in parallel ({parallel_failures}): {ul_err} ({err})"
+            ))
+          })?;
+
+          let wait_time = exponential_backoff(BASE_WAIT_TIME, i, MAX_WAIT_TIME);
+          sleep(Duration::from_millis(wait_time as u64)).await;
+
+          chunk = upload_chunk(&client, &url, &path, start, chunk_size).await;
+          i += 1;
+          drop(parallel_failure_permit);
+        }
+      }
+      drop(permit);
+      chunk.map_err(|e|{
+        //match e {
+        //  Error::Io(io) => Error::new_err(format!("Io error {io}")),
+        //  Error::Request(req) => Error::new_err(format!("Error while sending chunk {req}")),
+        //  Error::ToStrError(req) => Error::new_err(format!("Response header contains non ASCII chars: {req}")),
+        //}
+        e
+      }).map(|chunk| (part_number, chunk, chunk_size))
+    }));
+  }
+
+  let mut results: Vec<HashMap<String, String>> = vec![HashMap::default(); parts_urls.len()];
+
+  while let Some(result) = handles.next().await {
+    match result {
+      Ok(Ok((part_number, headers, size))) => {
+        //if let Some(ref callback) = callback {
+        //  callback.call((size,), None)?;
+        //}
+        results[part_number] = headers;
+      }
+      Ok(Err(py_err)) => {
+        return Err(py_err);
+      }
+      Err(err) => {
+        return Err(Error::new_err(format!(
+          "Error occurred while uploading: {err}"
+        )));
+      }
+    }
+  }
+
+  Ok(results)
+}
+
+async fn upload_chunk(
+  client: &reqwest::Client,
+  url: &str,
+  path: &str,
+  start: u64,
+  chunk_size: u64,
+) -> Result<HashMap<String, String>, Error> {
+  let mut options = OpenOptions::new();
+  let mut file = options.read(true).open(path).await?;
+  let file_size = file.metadata().await?.len();
+  let bytes_transferred = std::cmp::min(file_size - start, chunk_size);
+
+  file.seek(SeekFrom::Start(start)).await?;
+  let chunk = file.take(chunk_size);
+
+  let response = client
+    .put(url)
+    .header(CONTENT_LENGTH, bytes_transferred)
+    .body(reqwest::Body::wrap_stream(FramedRead::new(
+      chunk,
+      BytesCodec::new(),
+    )))
+    .send()
+    .await?;
+  let response = response.error_for_status()?;
+  let mut headers = HashMap::new();
+  for (name, value) in response.headers().into_iter() {
+    headers.insert(name.to_string(), value.to_str()?.to_owned());
+  }
+  Ok(headers)
+}
+

crates/desktop/tauri-4o/src/transfer/util.rs

@@ -0,0 +1,15 @@
+//! Implemented from https://github.com/huggingface/hf_transfer/blob/main/src/lib.rs
+
+use rand::{thread_rng, Rng};
+use std::fmt::Display;
+
+pub const BASE_WAIT_TIME: usize = 300;
+pub const MAX_WAIT_TIME: usize = 10_000;
+
+fn jitter() -> usize {
+  thread_rng().gen_range(0..=500)
+}
+
+pub fn exponential_backoff(base_wait_time: usize, n: usize, max: usize) -> usize {
+  (base_wait_time + n.pow(2) + jitter()).min(max)
+}